|
|
|
Including instrumentation and profiling |
Dynamic classification of sections of software6381735
Abstract
Dynamic classification of sections of software using a profile-based optimization system optimizes management of the sections of software. Software executes under expected usage conditions. After execution, a set of usage profiles describes the dynamic properties of sections of the software. Each usage profile includes information identifying a section of software. Each usage profile maps to an outcome meant to optimize management of the sections of the software during later execution. During such later execution, a usage background describes the dynamic properties of a section of the software. The usage background includes information identifying the section of software. By matching the usage background to a usage profile in the set of usage profiles, the section is dynamically classified during later execution. Based on this dynamic classification, the section maps to the outcome meant to optimize management of the sections of software.
Claims
I claim:
1. A method in a computer system for classifying a component of software by matching a usage background of the component to a usage profile, wherein the software comprises plural components, and wherein a dynamic structure tracks the usage of the plural components, the method comprising:
executing the software under expected conditions, thereby generating a usage profile, wherein executing comprises:
identifying a first component with a first component identifier; and
creating a usage profile based upon the first component identifier and the state of the dynamic structure;
mapping the usage profile to an outcome to produce a component map; and
re-executing the software, wherein re-executing comprises:
identifying a second component with a second component identifier;
creating a usage background of the second component based upon the second component identifier and the state of the dynamic structure;
matching the usage background to a usage profile in the component map; and
following the mapped outcome for the matched profile.
2. A computer-readable medium having computer-executable instructions for performing the method of claim 1.
3. The method of claim 1 further comprising:
during the executing, repeating the identifying and creating for each new first component, thereby generating plural usage profiles;
during the mapping, mapping the plural usage profiles to plural outcomes to produce additions to the component map; and
during the re-executing, repeating the identifying, creating, matching, and following for each new second component.
4. The method of claim 1 further comprising:
during the executing, repeating the identifying and creating, thereby generating a second usage profile;
comparing the second usage profile to the usage profile; and
if the second usage profile is not similar to the usage profile, mapping the second usage profile to an outcome to produce an addition to the component map.
5. The method of claim 1 further comprising:
re-executing the software under expected conditions, thereby generating a second usage profile;
comparing the second usage profile to the usage profile; and
if the second usage profile is not similar to the usage profile, mapping the second usage profile to an outcome to produce an addition to the component map.
6. The method of claim 1 wherein the plural components belong to plural groups, and wherein the first and second component identifiers are group identifiers.
7. The method of claim 1 wherein the first and second component identifiers are individual identifiers provided by a system service of the computer system.
8. The method of claim 1 wherein the first and second component identifiers are individual identifiers provided by an instrumentation system.
9. The method of claim 1 wherein for a component in use the dynamic structure records a component identifier, thereby tracking the sequence of usage of the plural components, wherein the creating a usage profile comprises:
noting a first sequence of recorded component identifiers in the dynamic structure; and
creating a first call chain with the first component identifier and the first sequence of recorded component identifiers;
and wherein the creating a usage background comprises:
noting a second sequence of recorded component identifiers in the dynamic structure; and
creating a second call chain with the second component identifier and the second sequence of recorded component identifiers.
10. The method of claim 1 wherein a component comprises plural functions, wherein for a function in use the dynamic structure records a function identifier, thereby tracking the sequence of usage of the plural functions of the plural components, wherein the creating a usage profile comprises:
noting a first sequence of recorded functions in the dynamic structure; and
creating a first call chain with the first component identifier and the first sequence of recorded function identifiers;
and wherein the creating a usage background comprises:
noting a second sequence of recorded functions in the dynamic structure; and
creating a second call chain with the second component identifier and the second sequence of recorded function identifiers.
11. The method of claim 1 wherein a component comprises plural functions, wherein for a function in use the dynamic structure records a function identifier and a component identifier of the component of which the function is a part, thereby tracking the sequence of usage of the plural components, wherein the creating a usage profile comprises:
noting a first sequence of recorded component identifiers and function identifiers in the dynamic structure; and
creating a first call chain with the first component identifier and the first sequence of recorded component identifiers and function identifiers;
and wherein the creating a usage background comprises:
noting a second sequence of recorded component identifiers and function identifiers in the dynamic structure; and
creating a second call chain with the second component identifier and the second sequence of recorded component identifiers and function identifiers.
12. A computer-readable medium having computer-executable instructions for performing the method of claim 11.
13. The method of claim 11 wherein the first and second call chains comprise for each use of a function the function identifier and the component identifier of the component of which the function is a part.
14. The method of claim 11 wherein the first and second call chains comprise for the first use of a component the component identifier and the function identifier of the function in use during the first use.
15. The method of claim 11 wherein a function identifier comprises a return address for a function.
16. The method of claim 1 wherein the dynamic structure is a system stack of the computer system.
17. The method of claim 1 wherein the dynamic structure is a stack provided by an instrumentation system.
18. The method of claim 1 wherein a component comprises plural functions, wherein the dynamic structure is a system stack of the computer system, wherein for a function in use the system stack records a function identifier, thereby tracking the sequence of usage of the plural functions, and wherein a second stack provided by an instrumentation system for a component in use records a component identifier, thereby tracking the sequence of usage of the plural components.
19. A computer-readable medium having computer-executable instructions for performing the method of claim 18.
20. The method of claim 1 wherein the creating a usage profile comprises:
setting a level of precision;
traversing the dynamic structure up to the level of precision; and
forming a usage profile based upon the first component identifier and the state of the traversed portion of the dynamic structure;
and wherein the creating a usage background comprises:
traversing the dynamic structure up to the level of precision; and
forming a usage background based upon the second component identifier and the state of the traversed portion of the dynamic structure.
21. The method of claim 20 wherein the dynamic structure is a stack, wherein the level of precision is a depth in the stack, wherein the traversing comprises following the stack to the set depth, and wherein the usage profile and usage background comprise call chains based upon the state of the traversed portion of the stack.
22. The method of claim 1 wherein the matching comprises:
finding the usage profile that is most similar to the usage background; and
selecting the mapped outcome for the most similar usage profile.
23. A method in a computer system for classifying a component of an application program by matching a usage background of the component to a usage profile, wherein the application comprises plural components, and wherein a dynamic structure tracks the usage of the plural components, the method comprising:
executing the application program in a profiling scenario, thereby generating a usage profile, wherein executing comprises:
identifying a first component with a first component identifier; and
creating a usage profile based upon the first component identifier and the state of the dynamic structure;
mapping the usage profile to a location in a distributed computing environment to produce a component map; and
re-executing the application program, wherein re-executing comprises:
identifying a second component with a second component identifier;
creating a usage background of the second component based upon the second component identifier and the state of the dynamic structure;
matching the usage background to a usage profile in the component map; and
instantiating the second component at the location that is mapped to the matched profile.
24. A computer-readable medium having computer-executable instructions for performing the method of claim 23.
25. The method of claim 23 further comprising:
during the executing, repeating the identifying and creating for each new first component, thereby generating plural usage profiles;
during the mapping, mapping the plural usage profiles to plural locations to produce additions to the component map; and
during the re-cxecuting, repeating the identifying, creating, matching, and instantiating for each new second component.
26. The method of claim 23 further comprising:
during the executing, repeating the identifying and creating, thereby generating a second usage profile;
comparing the second usage profile to the usage profile; and
if the second usage profile is not similar to the usage profile, mapping the second usage profile to a location to produce an addition to the component map.
27. The method of claim 23 further comprising:
re-executing the application program under expected conditions, thereby generating a second usage profile;
comparing the second usage profile to the usage profile; and
if the second usage profile is not similar to the usage profile, mapping the second usage profile to a location to produce an addition to the component map.
28. The method of claim 23 wherein a component belongs to a class, and wherein the first and second component identifier are class identifiers.
29. The method of claim 23 wherein the first and second component identifiers are individual identifiers provided by a system service of the computer system.
30. The method of claim 23 wherein the first and second component identifiers are individual identifiers provided by an instrumentation system.
31. The method of claim 23 wherein for an instantiated component the dynamic structure records a component identifier, thereby tracking the sequence of usage of the plural components, wherein the creating a usage profile comprises:
noting a first sequence of recorded component identifiers in the dynamic structure; and
creating a first call chain with the first component identifier and the first sequence of recorded component identifiers;
and wherein the creating a usage background comprises:
noting a second sequence of recorded component identifiers in the dynamic structure; and
creating a second call chain with the second component identifier and the second sequence of recorded component identifiers.
32. The method of claim 23 wherein a component is created by an activation function called by another component, and wherein the usage profile and usage background comprise the component identifier of a calling component.
33. The method of claim 23 wherein a component is created by an activation function called by another component, wherein the dynamic structure stores an identifier of the activation function, and wherein usage profile and usage background are based upon the identifier of the activation function.
34. The method of claim 23 wherein a component is created by an activation function called by another component, and wherein the usage profile and usage background comprise the component identifier of a calling component and an identifier of the called activation function.
35. The method of claim 23 wherein a component comprises plural functions, wherein for a function in use the dynamic structure records a function identifier, thereby tracking the sequence of usage of the plural functions of the plural components, wherein the creating a usage profile comprises:
noting a first sequence of recorded functions in the dynamic structure; and
creating a first call chain with the first component identifier and the first sequence of recorded function identifiers;
and wherein the creating a usage background comprises:
noting a second sequence of recorded functions in the dynamic structure; and
creating a second call chain with the second component identifier and the second sequence of recorded function identifiers.
36. The method of claim 23 wherein a component comprises plural functions, wherein for a function in use the dynamic structure records a function identifier and a component identifier of the component of which the function is a part, thereby tracking the sequence of usage of the plural components, wherein the creating a usage profile comprises:
noting a first sequence of recorded component identifiers and function identifiers in the dynamic structure; and
creating a first call chain with the first component identifier and the first sequence of recorded component identifiers and function identifiers;
and wherein the creating a usage background comprises:
noting a second sequence of recorded component identifiers and function identifiers in the dynamic structure; and
creating a second call chain with the second component identifier and the second sequence of recorded component identifiers and function identifiers.
37. A computer-readable medium having computer-executable instructions for performing the method of claim 36.
38. The method of claim 36 wherein the first and second call chains comprise for each use of a function the function identifier and the component identifier of the component of which the function is a part.
39. The method of claim 36 wherein the first and second call chains comprise for the first use of a component identifier and the function identifier of the function in use during the first use.
40. The method of claim 36 wherein a function identifier comprises a return address for a function.
41. The method of claim 23 wherein the dynamic structure is the call stack of the computer system.
42. The method of claim 23 wherein the dynamic structure is a stack provided by an instrumentation system.
43. A computer-readable medium having computer-executable instructions for performing the method of claim 42.
44. A computer-readable medium having computer-executable instructions for performing the method of claim 43.
45. The method of claim 23 wherein the creating a usage profile comprises:
setting a level of precision;
traversing the dynamic structure up to the level of precision; and
forming a usage profile based upon the first component identifier and the state of the traversed portion of the dynamic structure;
and wherein the creating a usage background comprises:
traversing the dynamic structure up to the level of precision; and
forming a usage profile based upon the second component identifier and the state of the traversed portion of the dynamic structure.
46. The method of claim 45 wherein the dynamic structure is a stack, wherein the level of precision is a depth in the stack, wherein the traversing comprises following the stack to the set depth, and wherein the usage profile and usage background comprise call chains based upon the state of the traversed portion of the stack.
47. The method of claim 23 wherein the matching comprises:
finding the usage profile that is most similar to the usage background; and
selecting the location that is mapped to the most similar usage profile.
48. A method in a computer system for classifying a program segment of software by matching a usage background of the program segment to a usage profile, wherein the software comprises plural program segments, and wherein a dynamic structure tracks the usage of the plural program segments, the method comprising:
executing the software under expected conditions, thereby generating a usage profile, wherein executing comprises:
identifying a first program segment with a first program segment identifier; and
creating a usage profile based upon the first program segment identifier and the state of the dynamic structure;
mapping the usage profile to an outcome to produce a program segment map, wherein the outcome is a memory placement; and
re-executing the software, wherein re-executing comprises:
identifying a second program segment with a second program segment identifier;
creating a usage background of the second program segment based upon the second program segment identifier and the state of the dynamic structure;
matching the usage background to a usage profile in the section map; and
following the mapped outcome for the matched profile, wherein the following comprises allocating memory for a program segment in the memory placement that is mapped to the matched usage profile.
49. The method of claim 48 further comprising:
during the executing, repeating the identifying and creating, thereby generating a second usage profile;
comparing the second usage profile to the usage profile; and
if the second usage profile is not similar to the usage profile, mapping the second usage profile to an outcome to produce an addition to the program segment map.
50. The method of claim 48 wherein the creating a usage profile comprises:
setting a level of precision;
traversing the dynamic structure up to the level of precision; and
forming a usage profile based upon the first program segment identifier and the state of the traversed portion of the dynamic structure;
and wherein the creating a usage background comprises:
traversing the dynamic structure up to the level of precision; and
forming a usage background based upon the second program segment identifier and the state of the traversed portion of the dynamic structure.
51. The method of claim 48 wherein the memory placement is a lifetime prediction bucket for garbage collection of memory.
52. The method of claim 48 wherein the memory placement is a memory heap for management of plural memory heaps.
53. A method in a computer system for classifying a program segment of software by matching a usage background of the program segment to a usage profile, wherein the software comprises plural program segments, and wherein a dynamic structure tracks the usage of the plural program segments, the method comprising:
executing the software under expected conditions, thereby generating a usage profile, wherein executing comprises:
identifying a first program segment with a first program segment identifier; and
creating a usage profile based upon the first program segment identifier and the state of the dynamic structure;
mapping the usage profile to an outcome to produce a program segment map, wherein the outcome is a lifetime estimate for load balancing; and
re-executing the software, wherein re-executing comprises:
identifying a second program segment with a second program segment identifier;
creating a usage background of the second program segment based upon the second program segment identifier and the state of the dynamic structure;
matching the usage background to a usage profile in the program segment map; and
following the mapped outcome for the matched profile, wherein the following comprises:
selecting a resource based on the lifetime estimate that is mapped to the matched usage profile; and
allocating memory for the second program segment in the selected resource.
54. The method of claim 53 further comprising:
during the executing, repeating the identifying and creating, thereby generating a second usage profile;
comparing the second usage profile to the usage profile; and
if the second usage profile is not similar to the usage profile, mapping the second usage profile to an outcome to produce an addition to the program segment map.
55. The method of claim 53 wherein for a program segment in use the dynamic structure records a program segment identifier, thereby tracking the sequence of usage of the plural program segments, wherein the creating a usage profile comprises:
noting a first sequence of recorded program segment identifiers in the dynamic structure; and
creating a first call chain with the first program segment identifier and the first sequence of recorded program segment identifiers;
and wherein the creating a usage background comprises:
noting a second sequence of recorded program segment identifiers in the dynamic structure; and
creating a second call chain with the second program segment identifier and the second sequence of recorded program segment identifiers.
56. The method of claim 53 wherein the creating a usage profile comprises:
setting a level of precision;
traversing the dynamic structure up to the level of precision; and
forming a usage profile based upon the first program segment identifier and the state of the traversed portion of the dynamic structure;
and wherein the creating a usage background comprises:
traversing the dynamic structure up to the level of precision; and
forming a usage background based upon the second program segment identifier and the state of the traversed portion of the dynamic structure.
57. The method of claim 56 wherein the dynamic structure is a stack, wherein the level of precision is a depth in the stack, wherein the traversing comprises following the stack to the set depth, and wherein the usage profile and usage background comprise call chains based upon the state of the traversed portion of the stack.
58. A method in a computer system for classifying a cache page of data in software by matching a usage background of the cache page of data to a usage profile, wherein the software comprises plural cache pages of data, and wherein a dynamic structure tracks the usage of the plural cache pages of data, the method comprising:
executing the software under expected conditions, thereby generating a usage profile, wherein executing comprises:
identifying a first cache page of data with a first cache page identifier; and
creating a usage profile based upon the first cache page identifier and the state of the dynamic structure;
mapping the usage profile to an outcome to produce a cache page map, wherein the outcome is a placement in a cache; and
re-executing the software, wherein re-executing comprises:
identifying a second cache page of data with a second cache page identifier;
creating a usage background of the second cache page of data based upon the second cache page identifier and the state of the dynamic structure;
matching the usage background to a usage profile in the cache page map; and
following the mapped outcome for the matched profile, wherein the following comprises writing a page of cache data to the cache based on the placement that is mapped to the matched usage profile.
59. The method of claim 58 further comprising:
during the executing, repeating the identifying and creating, thereby generating a second usage profile;
comparing the second usage profile to the usage profile; and
if the second usage profile is not similar to the usage profile, mapping the second usage profile to an outcome to produce an addition to the cache page map.
60. The method of claim 58 wherein the creating a usage profile comprises:
setting a level of precision;
traversing the dynamic structure up to the level of precision; and
forming a usage profile based upon the first cache page identifier and the state of the traversed portion of the dynamic structure;
and wherein the creating a usage background comprises:
traversing the dynamic structure up to the level of precision; and
forming a usage background based upon the second cache page identifier and the state of the traversed portion of the dynamic structure.
61. A method in a computer system for classifying an application component of software by matching a usage background of the application component to a usage profile, wherein the software comprises plural application components, and wherein a dynamic structure tracks the usage of the plural application components, the method comprising:
executing the software under expected conditions, thereby generating a usage profile, wherein executing comprises:
identifying a first application component with a first application component identifier; and
creating a usage profile based upon the first application component identifier and the state of the dynamic structure;
mapping the usage profile to an outcome to produce an application component map, wherein the outcome is a location in a distributed computing environment; and
re-executing the software, wherein re-executing comprises:
identifying a second application component with a second application component identifier;
creating a usage background of the second application component based upon the second application component identifier and the state of the dynamic structure;
matching the usage background to a usage profile in the application component map; and
following the mapped outcome for the matched profile, wherein the following comprises instantiating a component at the location that is mapped to the matched usage profile.
Description
TECHNICAL FIELD
The present invention relates generally to classification of sections of software by matching the usage background of the section to a usage profile derived by earlier profiling of the software.
BACKGROUND OF THE INVENTION
Fueled by the growing importance of the Internet, interest in the area of distributed systems (two or more computers connected by a communications medium) has increased in recent years. Programmers desiring to take advantage of distributed systems modify existing application programs to perform on distributed systems, or design applications for placement on distributed systems.
A distributed application is an application containing interconnected application units ("units") that are placed on more than one computer in a distributed system. By placing units on more than one computer in a distributed system, a distributed application can exploit the capabilities of the distributed system to share information and resources, and to increase application reliability and system extensibility. Further, a distributed application can efficiently utilize the varying resources of the computers in a distributed system.
Various types of modular software, including software designed in an object-oriented framework, can conceivably be distributed throughout a distributed system. Object-oriented programming models, such as the Microsoft Component Object Model ("COM"), define a standard structure of software objects that can be interconnected and collectively assembled into an application (which, being assembled from component objects, is herein referred to as a "component application"). The objects are hosted in an execution environment created by system services, such as the object execution environments provided by COM. This system exposes services for use by component application objects in the form of application programming interfaces ("APIs"), system-provided objects and system-defined object interfaces. Distributed object systems such as Microsoft Corporation's Distributed Component Object Model (DCOM) and the Object Management Group's Common Object Request Broker Architecture (CORBA) provide system services that support execution of distributed applications.
In accordance with object-oriented programming principles, the component application is a collection of object classes which each model real world or abstract items by combining data to represent the item's properties with functions to represent the item's functionality. More specifically, an object is an instance of a programmer-defined type referred to as a class, which exhibits the characteristics of data encapsulation, polymorphism and inheritance. Data encapsulation refers to the combining of data (also referred to as properties of an object) with methods that operate on the data (also referred to as member functions of an object) into a unitary software component (i.e., the object), such that the object hides its internal composition, structure and operation and exposes its functionality to client programs that utilize the object only through one or more interfaces. An interface of the object is a group of semantically related member functions of the object. In other words, the client programs do not access the object's data directly, but instead call functions on the object's interfaces to operate on the data. Polymorphism refers to the ability to view (i.e., interact with) two similar objects through a common interface, thereby eliminating the need to differentiate between two objects. Inheritance refers to the derivation of different classes of objects from a base class, where the derived classes inherit the properties and characteristics of the base class.
An application containing easily identifiable and separable units is more easily distributed throughout a distributed system. One way to identify separable units is to describe such units with structural metadata about the units. Metadata is data that describes other data. In this context, structural metadata is data describing the structure of application units. Further, application units are desirably location-transparent for in-process, cross-process, and cross-computer communications. In other words, it is desirable for communications between application units to abstract away location of application units. This flexibly enables the distribution of application units.
The partitioning and distribution of applications are problematic and complicated by many factors.
To partition an application for distribution, a programmer typically determines a plan for distributing units of the application based on past experience, intuition, or data gathered from a prototype application. The application's design is then tailored to the selected distribution plan. Even if the programmer selects a distribution plan that is optimal for a particular computer network, the present-day distribution plan might be rendered obsolete by changes in network topology. Moreover, assumptions used in choosing the distribution plan might later prove to be incorrect, resulting in an application poorly matched to its intended environment.
Generally, to distribute an application, one can work externally or internally relative to the application. External distribution mechanisms work without any modification of the application and include network file systems and remote windowing systems on a distributed system. Although external distribution mechanisms are easy to use and flexible, they often engender burdensome transfers of data between nodes of the distributed system, and for this reason are far from optimal. Internal distribution mechanisms typically modify the application to be distributed in various ways. Internal distribution mechanisms allow optimized application-specific distribution, but frequently entail an inordinate amount of extra programmer effort to find an improved distribution and modify the application. Further, internal systems frequently provide ad hoc, one-time results that are tied to the performance of a particular network at a particular time.
Automatic Distributed Partitioning Systems
An automatic distributed partitioning system (ADPS) works internally relative to an application to partition application units, and works automatically or semi-automatically to save programmer effort in designing distributed applications.
In the 1970's, researchers postulated that the best way to create a distributed application was to use a compiler in a run time environment to partition the application, and to provide the exact same code base to each of plural distributed machines as used on a single machine to execute the distributed application. After analyzing the structure of procedures and parameters in the source code of an application, metadata describing the structure of an application were generated from the application source code. Using this metadata, these ADPSs profiled the application and generated a communication model for the application. The Interconnected Processor System (ICOPS) is an example of an ADPS designed in the 1970's. The Configurable Applications for Graphics Employing Satellites (CAGES) also supported creation of distributed applications, but did not support automatic application profiling at all. A more recent example of an ADPS is the Intelligent Dynamic Application Partitioning (IDAP) System. ICOPS, CAGES, and IDAP suffer from numerous drawbacks relating to the universality, efficiency, and automation of these systems.
ICOPS, CAGES, and IDAP use static-type classification to partition an application and distribute it through a distributed computing environment. ICOPS and CAGES partition and distribute static functions and data such as application procedures using classification based upon the structure of the static functions and data. IDAP partitions and distributes statically allocated components using classification based upon the static type of the components. Static-type classification is inadequate for classifying application units that are dynamically instantiated and destroyed throughout execution, such as those designed according to COM and other object-oriented programming models. Deterioration of static-type classification is especially pronounced when application execution is data or input driven.
Lifetime Prediction and Garbage Collection Involving Dynamically Allocated Memory
Dynamically allocated memory is memory that is allocated to specific uses as needed, without specifying how much memory is needed in advance. When memory is no longer needed for the specific use, it is freed for later allocation to other uses. The process of freeing memory can be costly where it involves searching memory for program segments or data that are no longer active. A dynamically allocated object is an object for which space is allocated in memory as needed.
Dynamic classification has found limited application in the field of dynamically allocated memory. Lifetime prediction techniques dynamically classify an object to predict the behavior of the object. In memory management systems using memory heaps, a heap can become fragmented if long-lived and short-lived program segments are randomly allocated from the same heap. Memory heap fragmentation can be minimized if objects with short lifetimes are allocated memory from a separate heap. A process known as "garbage collection" identifies memory that is no longer needed, and frees such memory. Garbage collection is more efficient if dynamically allocated objects are grouped with objects having similar lifetimes. Lifetime prediction can reduce memory overhead and improve reference locality.
One technique for predicting lifetime of a dynamically allocated object uses a stack pointer at the time the object is dynamically allocated to identify the object. This technique is highly dependent on processor architecture. Another technique uses a hashed value of the return addresses from each of the invocation frames in a call stack to identify an object. Because this technique traverses the call stack accessing invocation frames, it is called a procedure call chain (PCC). The depth to which the call stack is traversed can vary. A depth-n PCC consists of the return addresses from each of the top n invocation frames. In general, the accuracy of a lifetime prediction increases with the depth of the PCC. This technique lacks sufficient contextual information to dynamically classify an object in an object-oriented system. PCCs form a sparse, one-dimensional space: the range of valid return addresses. Object-oriented programming adds a second dimension: the identity of the object executing the code. Dynamic memory allocation techniques limit contextual information in order to speed up the process of dynamic classification. None of these techniques for predicting lifetime of dynamically allocated objects utilizes available contextual information about the dynamically allocated objects themselves. Specifically, none utilizes identity of dynamically allocated objects for dynamic classification.
SUMMARY OF THE INVENTION
The present invention pertains to classification of a section of software by matching the usage background of the section to a usage profile determined by previous profiling software. As software executes, a dynamic structure tracks the usage of sections of the software and provides state information used to create the usage profile and usage background. Software executes under expected usage conditions, generating a usage profile. The usage profile includes information about the identity of a section being profiled and the usage of other sections of the software preceding the usage of the section being profiled. A section map includes a mapping of the usage profile to an outcome for the usage profile. When the software executes again, a usage background includes information about the identity of a section being classified and the usage of other section of the software leading up to the usage of a section being classified. In the section map, the usage background is matched to the usage profile most similar to it. The outcome that is mapped to the matched usage profile is followed.
According to an illustrated embodiment of the present invention, the sections of software are application units. The application executes in one or more profiling scenarios, generating usage profiles. These usage profiles map to locations in a distributed computing environment. During execution of the application outside of the profiling scenarios, a usage background includes information about an application unit and the preceding usage of application units. After matching the usage background to a usage profile in the unit map, a unit instantiates at the appropriate location in the distributed computing environment.
A usage profile can correspond to the usage of a single section of software in a single profiling scenario. Alternatively, a usage profile can correspond to the usage of multiple sections of software in a single profiling scenario, or the usage of multiple sections of software across multiple profiling scenarios.
The usage profile and usage background include information about the identity of a section of software being profiled or classified and the preceding usage of other sections. The identity information is an identifier for a group to which the section belongs. Alternatively, the identity information is an identifier for an individual section assigned by an operating system or by an instrumentation system.
Information about the preceding usage of other sections includes identity information for preceding sections. For example, a call chain records the identities of sections called leading up to the section being profiled or classified. Alternatively, information about the preceding usage of other sections can include identifiers of functions used by preceding sections. For example, instead of or in addition to the section identity call chain, a call chain can record the return addresses of functions used leading up to the section being profiled or classified. A call chain can include various subsets of section identities and function identifiers to enable varying levels of precision in dynamic classification. For example, a call chain can include only an identifier from the section immediately preceding the section being classified. In the illustrated embodiment, a usage profile and usage background can be a table entry including the identifier of a unit creating the unit being profiled or classified. Alternatively, the table entry can include a description of the function called to create the unit being profiled or classified.
A dynamic structure tracks the usage of sections preceding the section being profiled or classified. The dynamic structure tracks sections and/or functions used before the section being classified. Instead of a single dynamic structure, multiple dynamic structures can be used. For example, in the illustrated embodiment, a call stack tracks function calls while a separate stack tracks unit identities.
The outcome that is mapped to a usage profile corresponds to an event that occurs when a usage background matches a usage profile. In the illustrated embodiment, an outcome is a location in a distributed computing environment for an application unit to be instantiated. Alternatively, the dynamic classification techniques of the present invention can be applied to lifetime prediction for memory heap management, garbage collection, load balancing among resources, cache management, or other applications in which sections of software are profiled then classified to optimize performance.
Additional features and advantages of the present invention will be made apparent from the following detailed description of an illustrated embodiment, which proceeds with reference to the accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a diagram of a distributed computing environment in which the present invention can be implemented.
FIG. 2 is a block diagram of a computer system that can be used to implement the present invention.
FIG. 3 is a block diagram of a Microsoft Component Object Model software component that can be used to implement the present invention.
FIG. 4 is a block diagram of a client and the component of FIG. 3 in a distributed computing environment.
FIG. 5 is a block diagram of the component of FIG. 3 with multiple interfaces specified according to Microsoft's Component Object Model.
FIG. 6 is a flow chart showing the automatic partitioning of an application into application units according to the illustrated embodiment of the present invention.
FIG. 7 is a flow chart showing the scenario-based profiling of an application to generate a description of the run-time behavior of the application according the illustrated embodiment of the present invention.
FIG. 8 is a commodity flow diagram cut by a MIN CUT MAX FLOW algorithm according to the illustrated embodiment of the present invention.
FIG. 9 is a listing showing a code fragment in which a component like that illustrated in FIG. 3 is created, and types of dynamic classifiers for the component.
FIG. 10 is a listing containing code fragments illustrating various techniques for intercepting communications according to the illustrated embodiment of the present invention.
FIG. 11 is a diagram showing a graphical representation of a distribution chosen for a profiled scenario in which the user loads and previews an image in Picture It!.RTM. from a server in the COIGN system.
FIG. 12 is a block diagram of an object-oriented framework for partitioning and distributing application units of an application according to the COIGN system.
FIG. 13 is a block diagram of an object-oriented framework for partitioning and distributing application units of an application showing the pattern of intercommunication between the objects according to the COIGN system.
FIG. 14 is a listing containing code fragments illustrating interception and in-line redirection of communications according to the COIGN system.
FIG. 15 is a block diagram showing an application binary in common object file format that is statically linked according to one embodiment of the present invention.
FIG. 16 is a block diagram showing the application binary of FIG. 15 reversibly static re-linked to a second set of libraries.
FIG. 17 is a block diagram of a series of COIGN data structures showing a component object, an interface wrapper appended to the component object, and analytical data appended to the wrapped component object.
FIG. 18 is a block diagram of a series of COIGN data structures showing a table of interfaces, a group of interface wrappers, and a table of instrumentation functions.
DETAILED DESCRIPTION OF AN ILLUSTRATED EMBODIMENT
The present invention is directed toward automatic partitioning of units of an application and distribution of those units. In the illustrated embodiment of the present invention, an application is partitioned into one or more application units for distribution in a distributed computing environment. The COIGN system is one possible refinement of the illustrated ADPS that automatically partitions and distributes applications designed according to the Component Object Model ("COM") of Microsoft Corporation of Redmond, Washington. Briefly described, the COIGN system includes techniques for identifying COM components, measuring communication between COM components, classifying COM components, measuring network behavior, detecting component location constraints, generating optimal distribution schemes, and distributing COM components during run-time.
FIGS. 1 and 2 and the following discussion are intended to provide a brief, general description of a suitable computing environment in which the illustrated ADPS can be implemented. While the present is described in the general context of computer-executable instructions that run on computers, those skilled in the art will recognize that the present invention can be implemented as a combination of program modules, or in combination with other program modules. Generally, program modules include routines, programs, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The present invention can be implemented as a distributed application, one including program modules located on different computers in a distributed computing environment.
Exemplary Distributed Computing Environment
FIG. 1 illustrates a distributed computing environment 1 in which units of an application are partitioned and distributed by the illustrated ADPS in accordance with the present invention. The distributed computing environment 1 includes two computer systems 5 connected by a connection medium 10. The computer systems 5 can be any of several types of computer system configurations, including personal computers, hand-held devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, minicomputers, mainframe computers, and the like. In terms of logical relation with other computer systems 5, a computer system 5 can be a client, a server, a router, a peer device, or other common network node. Moreover, although FIG. 1 illustrates two computer systems 5, the present invention is equally applicable to an arbitrary, larger number of computer systems connected by the connection medium 10. Further, the distributed computing environment 1 can contain an arbitrary number of additional computer systems 5 which do not directly involve the illustrated ADPS, connected by an arbitrary number of connection mediums 10. The connection medium 10 can comprise any local area network (LAN), wide area network (WAN), or other computer network, including but not limited to Ethernets, enterprise-wide computer networks, intranets and the Internet.
The illustrated ADPS automatically partitions an application and distributes program units by locating them in more than one computer system 5 in the distributed computing environment 1. Portions of the illustrated ADPS can be implemented in a single computer system 5, with the application later distributed to other computer systems 5 in the distributed computing environment 1. Portions of the illustrated ADPS can also be practiced in a distributed computing environment 1 where tasks are performed by a single computer system 5 acting as a remote processing device that is accessed through a communications network, with the distributed application later distributed to other computer systems 5 in the distributed computing environment 1. In a networked environment, program modules of the illustrated ADPS can be located on more than one computer system 5.
Exemplary Computer System
FIG. 2 illustrates an example of a computer system 5 that can serve as an operating environment for the illustrated ADPS. With reference to FIG. 2, an exemplary computer system for implementing the invention includes a computer 20 (such as a personal computer, laptop, palmtop, set-top, server, mainframe, and other varieties of computer), including a processing unit 21, a system memory 22, and a system bus 23 that couples various system components including the system memory to the processing unit 21. The processing unit can be any of various commercially available processors, including Intel x86, Pentium and compatible microprocessors from Intel and others, including Cyrix, AMD and Nexgen; Alpha from Digital; MIPS from MIPS Technology, NEC, IDT, Siemens, and others; and the PowerPC from IBM and Motorola. Dual microprocessors and other multi-processor architectures also can be used as the processing unit 21.
The system bus can be any of several types of bus structure including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of conventional bus architectures such as PCI, VESA, AGP, Microchannel, ISA and EISA, to name a few. The system memory includes read only memory (ROM) 24 and random access memory (RAM) 25. A basic input/output system (BIOS), containing the basic routines that help to transfer information between elements within the computer 20, such as during start-up, is stored in ROM 24.
The computer 20 further includes a hard disk drive 27, a magnetic disk drive 28, e.g., to read from or write to a removable disk 29, and an optical disk drive 30, e.g., for reading a CD-ROM disk 31 or to read from or write to other optical media. The hard disk drive 27, magnetic disk drive 28, and optical disk drive 30 are connected to the system bus 23 by a hard disk drive interface 32, a magnetic disk drive interface 33, and an optical drive interface 34, respectively. The drives and their associated computer-readable media provide nonvolatile storage of data, data structures, computer-executable instructions, etc. for the computer 20. Although the description of computer-readable media above refers to a hard disk, a removable magnetic disk and a CD, it should be appreciated by those skilled in the art that other types of media which are readable by a computer, such as magnetic cassettes, flash memory cards, digital video disks, Bernoulli cartridges, and the like, can also be used in the exemplary operating environment.
A number of program modules can be stored in the drives and RAM 25, including an operating system 35, one or more application programs 36, other program modules 37, and program data 38.
A user can enter commands and information into the computer 20 through a keyboard 40 and pointing device, such as a mouse 42. Other input devices (not shown) can include a microphone, joystick, game pad, satellite dish, scanner, or the like. These and other input devices are often connected to the processing unit 21 through a serial port interface 46 that is coupled to the system bus, but can be connected by other interfaces, such as a parallel port, game port or a universal serial bus (USB). A monitor 47 or other type of display device is also connected to the system bus 23 via an interface, such as a video adapter 48. In addition to the monitor, computers typically include other peripheral output devices (not shown), such as speakers and printers.
The computer 20 can operate in a networked environment using logical connections to one or more other computer systems 5. The other computer systems 5 can be servers, routers, peer devices or other common network nodes, and typically include many or all of the elements described relative to the computer 20, although only a memory storage device 49 has been illustrated in FIG. 2. The logical connections depicted in FIG. 2 include a local area network (LAN) 51 and a wide area network (WAN) 52. Such networking environments are commonplace in offices, enterprise-wide computer networks, intranets and the Internet.
When used in a LAN networking environment, the computer 20 is connected to the local network 51 through a network interface or adapter 53. When used in a WAN networking environment, the computer 20 typically includes a modem 54 or other means for establishing communications (e.g., via the LAN 51 and a gateway or proxy server 55) over the wide area network 52, such as the Internet. The modem 54, which can be internal or external, is connected to the system bus 23 via the serial port interface 46. In a networked environment, program modules depicted relative to the computer 20, or portions thereof, can be stored in the remote memory storage device. It will be appreciated that the network connections shown are exemplary and other means of establishing a communications link between the computer systems 5 (including an Ethernet card, ISDN terminal adapter, ADSL modem, 10BaseT adapter, 100BaseT adapter, ATM adapter, or the like) can be used.
In accordance with the practices of persons skilled in the art of computer programming, the illustrated ADPS is described below with reference to acts and symbolic representations of operations that are performed by the computer 20, unless indicated otherwise. Such acts and operations are sometimes referred to as being computer-executed. It will be appreciated that the acts and symbolically represented operations include the manipulation by the processing unit 21 of electrical signals representing data bits which causes a resulting transformation or reduction of the electrical signal representation, and the maintenance of data bits at memory locations in the memory system (including the system memory 22, hard drive 27, floppy disks 29, and CD-ROM 31) to thereby reconfigure or otherwise alter the computer system's operation, as well as other processing of signals. The memory locations where data bits are maintained are physical locations that have particular electrical, magnetic, or optical properties corresponding to the data bits.
Component Object Overview
With reference now to FIG. 3, in the COIGN system, the computer 20 (FIG. 2) executes "COIGN," a component-based application that is developed as a package of component objects. COIGN's component objects conform to the Microsoft Component Object Model ("COM") specification (i.e., each is implemented as a "COM Object" 60, alternatively termed a "COM component"). COIGN executes using the COM family of services (COM, Distributed COM ("DCOM"), COM+) of the Microsoft Windows NT Server operating system, but alternatively can be implemented according to other object standards (including the CORBA (Common Object Request Broker Architecture) specification of the Object Management Group) and executed under object services of another operating system.
COIGN automatically partitions and distributes other component-based applications. Like COIGN, the component-based applications automatically partitioned and distributed by COIGN are implemented in conformity with COM and executed using COM services, but alternatively can be implemented according to another object standard and executed using object services of another operating system.
COM: Binary Compatibility
The COM specification defines binary standards for objects and their interfaces which facilitate the integration of software components into applications. COM specifies a plafform-standard binary mapping for interfaces, but does not specify implementations for interfaces. In other words, an interface is defined, but the implementation of the interface is left up to the developer. The binary format for a COM interface is similar to the common format of a C++ virtual function table. Referring to FIG. 3, in accordance with COM, the COM object 60 is represented in the computer system 20 (FIG. 2) by an instance data structure 62, a virtual function table 64, and member methods (also called member functions) 66-68. The instance data structure 62 contains a pointer 70 to the virtual function table 64 and data 72 (also referred to as data members, or properties of the object). A pointer is a data value that holds the address of an item in memory. The virtual function table 64 contains entries 76-78 for the member methods 66-68. Each of the entries 76-78 contains a reference to the code 66-68 that implements the corresponding member methods. A reference to an interface is stored as a pointer to the pointer 70.
While extremely simple, the binary mapping provides complete binary compatibility between COM components written in any language with any development tool. Any language that can call a function through a pointer can use COM components. Any language that can export a function pointer can create COM components. Language-neutral binary compatibility is an important feature of COM.
COM: Strongly Typed Interfaces and Interface Descriptor Language
The pointer 70, the virtual function table 64, and the member methods 66-68 implement an interface of the COM object 60. By convention, the interfaces of a COM object are illustrated graphically as a plug-in jack as shown in objects 110 and 130 in FIG. 4. Also, interfaces conventionally are given names beginning with a capital "I." In accordance with COM, the COM object 60 can include multiple interfaces, which are implemented with one or more virtual function tables. The member function of an interface is denoted as "llnterfaceName::MethodName."
All first-class communication in COM takes place through well-defined, binary-standard interfaces, which are strongly typed references to a collection of semantically related functions.
Programmatically, interfaces are described either with an Interface Definition Language (IDL) or with a package of compiled metadata structures called a type library. Whether expressed in IDL or a type library, the interface definition enumerates in detail the number and type of all arguments passed through interface functions. Each interface function can have any number of parameters. To clarify semantic features of the interface, IDL attributes can be attached to each interface, member function, or parameter. In IDL syntax, attributes are enclosed in square brackets ([]). Attributes specify features such as the data-flow direction of function arguments, the size of dynamic arrays, and the scope of pointers. Syntactically, IDL is very similar to C++. Moreover, the interface definition has a purpose similar to that of a function prototype in C++; it provides a description for invocation, but not an implementation. An IDL compiler maps the interface definitions into a standard format for languages such as C++, Java, or Visual Basic. For example, the Microsoft IDL compiler, MIDL, can map interfaces into C++ or export compiled IDL metadata to a type library. (For a detailed discussion of COM and OLE, see Kraig Brockschmidt, Inside OLE, Second Edition, Microsoft Press, Redmond, Wash. (1995)).
COM: Globally Unique Identifiers
In COM, classes of COM objects are uniquely associated with class identifiers ("CLSIDs"), and registered by their CLSID in the registry. The registry entry for a COM object class associates the CLSID of the class with information identifying an executable file that provides the class (e.g., a DLL file having a class factory to produce an instance of the class). Class identifiers are 128-bit globally unique identifiers ("GUIDs") that the programmer creates with a COM service named "CoCreateGUID" (or any of several other APIs and utilities that are used to create universally unique identifiers) and assigns to the respective classes. The interfaces of a component are also immutably associated with interface identifiers ("IIDs"), which are also 128-bit GUIDs. If an interface changes, it receives a new IID.
COM: Implementation
The virtual function table 64 and member methods 66-68 of the COM object 60 are provided by an object server program 80 (hereafter "object server DLL") which is stored in the computer 20 (FIG. 2) as a dynamic link library file (denoted with a ".dll" file name extension). In accordance with COM, the object server DLL 80 includes code for the virtual function table 64 and member methods 66-68 of the classes that it supports, and also includes a class factory 82 that generates the instance data structure 62 for an object of the class.
Other objects and programs (referred to as a "client" of the COM object 60) access the functionality of the COM object by invoking the member methods through the COM object's interfaces. First, however, the COM object must be instantiated (i.e., by causing the class factory to create the instance data structure 62 of the object); and the client must obtain an interface pointer to the COM object.
Before the COM object 60 can be instantiated, the object is first installed on the computer 20. Typically, installation involves installing a group of related objects called a package. The COM object 60 is installed by storing the object server DLL file(s) 80 that provides the object in data storage accessible by the computer 20 (typically the hard drive 27, shown in FIG. 2), and registering COM attributes (e.g., class identifier, path and name of the object server DLL file 80, etc.) of the COM object in the system registry. The system registry is a per-machine component configuration database.
COM: Component Instantiation
A client requests instantiation of the COM object locally or on a remote computer using system-provided services and a set of standard, system-defined component interfaces based on class and interface identifiers assigned to the COM Object's class and interfaces. More specifically, the services are available to client programs as application programming interface (API) functions provided in the COM library, which is a component of the Microsoft Windows NT operating system in a file named "OLE32.DLL." The DCOM library, also a component of the Microsoft Windows NT operating system in "OLE32.DLL," provides services to instantiate COM objects remotely and to transparently support communication among COM objects on different computers.
In particular, the COM library provides "activation mechanism" API functions, such as "CoCreateInstanceo( )" that the client program can call to request local or remote creation of a component using its assigned CLSID and an IID of a desired interface. In response to a request, the "CoCreateInstance( )" API looks up the registry entry of the requested CLSID in the registry to identify the executable file for the class. The "CoCreateInstance( )" API function then loads the class' executable file either in the client program's process, or into a server process which can be either local or remote (i.e., on the same computer or on a remote computer in a distributed computer network) depending on the attributes registered for the COM object 60 in the system registry. The "CoCreateInstance( )" API uses the class factory in the executable file to create an instance of the COM object 60. Finally, the "CoCreateInstance( )" API function returns a pointer of the requested interface to the client program.
Referring to FIG. 4, a system including a local client 100 and a remote component 140 is described. A local client 100 instantiates and accesses the services of a remote component 140 using services provided by DCOM. DCOM provides the low-level services supporting instantiation of component 140 in another process or on another machine. After instantiation, DCOM supports cross-process or cross-machine communication.
More specifically, after the "CoCreatInstance" API 102 of the OLE32 DLL 104 is called by a client 100, the "CoCreatInstance" API 102 determines from the system registry, from an explicit parameter, or from a moniker, the class of the component 140 and in which machine or process the component 140 should be instantiated. In FIG. 4, the component 140 is to be activated 106 on a remote machine. A local Service Control Manager 108 connects to a remote Service Control Manager 144, which requests creation of the component 140 through the "CoCreatInstance" API 102. An executable file 80 for the class is then loaded into a remote server process, and the class factory 82 in the executable file 80 is used to create an instance of the COM object 140. Finally, the "CoCreatInstance( )" API 102 function returns to the client 100 an interface pointer to an interface proxy 110 for the requested component 140. Whether a component is instantiated locally or remotely, the pointer returned to the client program refers to a location in local address space. So to a client, all component instantiations appear to be in-process.
COM: In-Process, Cross-Process, and Cross-Machine Communication
Binary compatibility gives COM components true location transparency. A client can communicate with a COM component in the same process, in a different process, or on an entirely different machine. Stated more succinctly, COM supports in-process, cross-process, or cross-machine communication. The location of the COM component is completely transparent to the client because in each case the client still invokes the component by calling indirectly through an interface's virtual function table. Location transparency is supported by two facilities: MIDL generation of interface proxies and stubs, and the system registry.
Referring again to FIG. 4, cross-machine communication occurs transparently through and interface proxy 110 and stub 130, which are generated by software such as the MIDL compiler. The proxy 110 and stub 130 include information necessary to parse and type function arguments passed between the client 100 and the component 140. For example, this information can be generated from an Interface Description Language (IDL) description of the interface of the component 140 that is accessed by the client 100. The proxy 110 and stub 130 can provide security for communication between the client 100 and the component 140. A client 100 communicates with the proxy 110 as if the proxy 110 were the instantiated component 140. The component 140 communicates with the stub 130 as if the stub 130 were the requesting client 100. The proxy 110 marshals function arguments passed from the client into one or more packets that can be transported between address spaces or between machines. Data for the function arguments is stored in a data representation understood by both the proxy 110 and the stub 130. In DCOM, the proxy 110 and stub 130 copy pointer-rich data structures using deep-copy semantics. The proxy 110 and stub 130 typically include a protocol stack and protocol information for remote communication, for example, the DCOM network protocol, which is a superset of the Open Group's Distributed Computing Environment Remote Procedure Call (DCE RPC) protocol. The one or more serialized packets are sent over the network 120 to the destination machine. The stub unmarshals the one or more packets into function arguments, and passes the arguments to the component 140. In theory, proxies and stubs come in pairs--the first for marshaling and the second for unmarshaling. In practice, COM combines code for the proxy and stub for a specific interface into a single reusable binary.
The client 100 invokes the component 140 through an indirect call on an interface virtual function table 64. In this case, however, following the interface pointer provided to the client 100, the virtual function table 64 belongs to the proxy 110. The proxy 110 marshals function argument into one or more serialized packets and sends the packets to the destination machine using DCOM Network Protocol. The stub 130 unmarshals the arguments and calls the component 140 through the interface virtual function table 64 in the target address space. As a call is returned, the process is reversed. In this way, in-process communication between client 100 and component 140 is emulated in a distributed computing environment, invisibly to both the client 100 and the component 140.
Invocation of cross-process components is very similar to invocation of cross-machine components. Moreover, cross-process communication uses the same interface proxies and stubs as cross-machine communication. The important difference is that once the function arguments have been marshaled into a buffer, DCOM transfers execution to the address space of the component. As with cross-machine invocation and communication, cross-process invocation and communication are completely transparent to both client and component.
COM insures location transparency because all communication takes place through calls on interface virtual function tables. The client does not know whether the code pointed to by the virtual function table belongs to the component or to an interface proxy that will forward the message to the remote component.
COM: Standard Interfaces
Once the client of the COM object 60 has obtained the first interface pointer of the COM object, the client can obtain pointers of other desired interfaces of the component using the interface identifier associated with the desired interface.
The "IUnknown" interface includes a member function named "Queryinterface( )." The "QueryInterface( )" function can be called with an interface identifier as an argument, and returns a pointer to the interface associated with that interface identifier. The "IUnknown" interface of each COM object also includes member functions, "AddRef( )" and "Release( )." Whenever a client of a component creates a new reference (e.g., an interface pointer) to the component, it calls "AddRe( )." When it is finished using the reference, it calls "Release( )." Through the "AddRef( )" and "Release( )" functions, a component knows exactly how many clients have references to it. When its reference count goes to zero, the component is responsible for freeing itself from memory. By convention, the "IUnknown" interface's member functions are included as part of each interface on a COM object. Thus, any interface pointer that the client obtains to an interface of a COM object can be used to call the "Queryinterface( )" function.
Com: Interface Design Considerations
By design, the COM binary standard restricts the implementation of an interface and components to the degree necessary to insure interoperability. To summarize, COM places four specific restrictions on interface design to insure component interoperability. First, a client accesses a component through its interface pointers. Second, the first item pointed to by an interface pointer must be a pointer to a virtual function table. Third, the first three entries of the virtual function table must point to the "Queryinterface( )", "AddRef( )" and "Release( )" functions for the interface. Finally, if a client intends to use an interface, it must insure that the interface's reference count has been incremented. As long as a component programmer obeys the four rules of the COM binary standard, he or she is completely free to make any other implementation choices.
During implementation, the component programmer chooses a memory layout for component and per-instance interface data. Memory layout is influenced by the number of supported interfaces, the existence of unique instances of the same interface for different clients, the expected lifetimes of interface instances, the amount of per-instance and per-component data, and internal, component-specific design factors.
Most components support at most roughly a dozen interfaces with each interface having only a single instance. Referring to FIG. 5, the relationship between a client 100 and a component 140 exposing multiple interfaces to the client is explored in some detail. The client includes an interface pointer 160 to the IUnknown interface, and other interface pointers 162-166 for other interfaces exposed by the client. The interface pointers 160-166 point to an instance data structure 62 for the component 140. COM defines several standard interfaces generally supported by COM objects including the "lunknown" interface. A pointer 170 to the virtual table 180 is listed first in the instance data structure 62 of the component 140. The instance data structure 62 contains one VTBL pointer 170-173 per interface, a per-component reference count 176, and internal component data 178. Each VTBL pointer 170-173 points to a virtual table 180-183, which in turn contain pointers to member functions 190-195 of the interfaces. Every interface includes the "Queryinterface( )" 190, "AddRef( )" 191, and "Release( )" 192 functions. In addition, interfaces can include other member functions. For example, Interface3 includes the additional functions 193-195. Within the component's member functions, a constant value is added to the "this" pointer to find the start of the memory block and to access component data 178. All of the component interfaces use a common pair of "AddRef( )" and "Release( )" functions to increment and decrement the component reference count 176.
Sometimes, a component supports multiple copies of a single interface. Multiple-instance interfaces are often used for iteration. A new instance of the interface is allocated for each client. Multiple-instance interfaces are typically implemented using a tear-off interface. A tear-off interface is allocated as a separate memory block. The tear-off interface contains the interface's VTBL pointer, a per-interface reference count, a pointer to the component's primary memory block, and any instance-specific data. In addition to multiple-instance interfaces, tear-off interfaces are often used to implement rarely accessed interfaces when component memory size is desirably minimized, (i.e., when the cost of the extra four bytes for a VTBL pointer per component instance is too expensive).
Components commonly use a technique called delegation to export interfaces from another component to a client. Delegation is often used when one component aggregates services from several other components into a single entity. The aggregating component exports its own interfaces, which delegate their implementation to the aggregated components. In the simple case, the delegating interface simply calls the aggregated interface. The simple case is interface specific, code intensive, and requires an extra procedure call during invocation. The simple solution is code intensive because delegating code is written for each interface type. The extra procedure call becomes particularly important if the member function has a large number of arguments or multiple delegators are nested through layers of aggregation.
A generalization of delegation is the use of a universal delegator. The universal delegator is essentially a type-independent, re-usable delegator. The data structure for a universal delegator consists of a VTBL pointer, a reference count, a pointer to the aggregated interface, and a pointer to the aggregating component. Upon invocation, a member function in the universal delegator replaces the "this" pointer on the argument stack with the pointer to the delegated interface and jumps directly to the entry point of the appropriate member function in the aggregated interface. The universal delegator is "universal" because its member functions need know nothing about the type of interface to which they are delegating; they reuse the invoking call frame. Implemented in a manner similar to tear-off interfaces, universal delegators are instantiated on demand, one per delegated interface with a common VTBL shared among all instances.
Alternative Object Standards
Although COIGN is described with reference to applications designed according to COM, aspects of COIGN are equally applicable to applications designed according to other object standards. For example, the following aspects, later described in detail, are equally applicable to COM and non-COM applications: automatic distributed partitioning of an application binary; recording summarized pair-wise component communication; deriving a network-independent representation of application communication; re-instrumenting an application for distribution using pre-processed metadata; reversible static linking of a library to an application; in-line redirection of object creation requests in an ADPS; dynamic classification; quickly estimating network latency and bandwidth; and automatically detecting location constraints.
Alternative Distributed Communications Services
The COIGN system is described with reference to communication support provided by the COM family of services. Other distributed communication services provide cross-process and cross-machine transparency, but not in-process location transparency. This prevents a server process from running in the same address space as a client process, and thus prevents a distributed application from using inexpensive in-process communication between components also capable of distributed communication. In contrast, the COM family of services provides true location transparency, so non-distributed components pay no performance penalty for exposing potentially distributable interfaces.
Even so, a true location-transparent component system similar to COM could be built with some effort upon other distribution services, as in fact COM builds on the Distributed Computing Environment Remote Procedure Call ("DCERPC") standard. The COIGN system could then be ported to the new system.
Overview of the Illustrated ADPS
It is both possible and beneficial to partition and distribute applications automatically. Quantitatively, the benefit of automatic distributed partitioning is determined by the performance of the chosen distribution. It is possible to determine a distribution for a given application that minimizes communication costs for the application in a given distributed computing environment. Ultimately, however, the performance of a selected application distribution also depends on the granularity and quality of the application's units (e.g., COM objects in the COIGN system ADPS), and, where applicable, on the appropriateness of the profiling scenarios (described below) used to measure internal application communication. While the present invention cannot improve a completed application's design, it can achieve the best possible distribution of that design subject to the profiling scenarios.
Automatic distributed partitioning reduces the programmer's burden. Rather than code for a specific distribution, the programmer is encouraged to create easily distributed application units. Emphasis is placed on code reusability, application unit autonomy, and choice of appropriate algorithm and data abstractions--all elements of good software engineering. In essence, automatic distributed partitioning makes the most of good software engineering by raising the level of abstraction for the distributed application programmer. In contrast, manual distributed partitioning forces the programmer to be keenly aware of how an application will be distributed.
Distributed partitioning is complicated by interactions between code modules, between data structures, and between both code and data. For instance, one data structure can contain a pointer to another data structure. If either data structure is naively relocated to another machine without modification, an attempt to de-reference the pointer will fail, most likely producing a virtual memory fault. Automatic distributed partitioning requires that either the programmer or the computer system explicitly manage code and data interactions crossing machine boundaries. For example, in the COIGN system, the COM family of services manages code and data interactions across machine and process boundaries.
In general, an ADPS takes an application as its input. For output, the ADPS modifies the application to produce a distributed version of the application that minimizes network communication costs.
Referring to FIG. 6, an application 200 is automatically partitioned for distribution according to the illustrated embodiment of the present invention. In the illustrated ADPS, the application 200 is of design known in the art. In the COIGN system, for example, the application 200 is an application binary, including executable files, dynamic link libraries, and other object code representations of software. In the COIGN system, the application binary is desirably designed according to an object model with suitable granularity, location transparency, and interface description, for example, Microsoft's COM, but alternatively can be designed according to other standards.
An application description set 220 describing the behavior of the application is prepared at step 210 for the application 200. The application description set 220 can be supplied by an external source that analyzes the application 200 in advance, or can be generated by the illustrated ADPS itself. The application description set 220 can include static and/or dynamic metadata describing the application. For example, in the COIGN system, the application description set 220 can include static metadata derived from metadata provided by a Microsoft IDL compiler (MIDL). Alternatively, the application description set 220 can include static metadata generated by the illustrated ADPS through static analysis techniques. Dynamic analysis techniques can be used by the illustrated ADPS to include dynamic metadata (such as dynamic descriptions of units, descriptions of actual inter-unit communication between the units of the application 200, and descriptions of how much time was spent in each unit in computation) in the application description set 220.
An environment description set 230 describes the distributed computing environment in which the application 200 is to be distributed. The environment description set 230 can be a description of an idealized computer network with identical computers and no communication costs. Alternatively, the environment description set 230 includes a high level description of a particular physical network on which the application 200 is to be distributed. The environment description set 230 can include a high level behavioral classification scheme used to determine which units should run on particular machines in a distributed computing environment. The environment description set 230 can also include descriptions of network characteristics such as latency and bandwidth, or descriptions of location constraints for particular units. In an alternative embodiment, the application description set 220 implicitly contains description of the behavior of a distributed computing environment along with description of the behavior of an application, for example real-time measurements of communications between distributed units of an application.
The environment description set 230 and application description set 220 are analyzed at step 240 to determine where units of the application 200 should be located in the distributed computing environment, for example according to the following pseudocode:
If (unit behavior=x) locate unit on machine Y
Else locate unit on machine Z.
In the COIGN system, a more complicated algorithm, for example, a commodity flow algorithm, is applied to a representation of units and communication between the units.
A distribution scheme 50 is the result of applying the environment description set 230 to the application description set 220. The distribution scheme 250 includes a mapping of application units to locations in a distributed computing environment. The units can be classified using static metadata of the units. Alternatively, where run-time profiling was used to dynamically describe the units, the units can be classified according to dynamic behavior. At run-time, units of the application 200 are mapped using the distribution scheme 250 for location on an appropriate computer in the distributed computing environment.
The various aspects of the present invention can be organized according to the three sub-areas they involve: discovering how the application can be partitioned, deciding how the application should be distributed, and achieving a chosen distribution.
Discovery: Discovering How the Application Can be Partitioned
An application description set 220 describes the behavior of the application. In the illustrated ADPS, these descriptors can be supplied by an external source and include static and/or dynamic metadata about the application. In the COIGN system, COIGN generates the application description set using an instrumentation package attached to the application, identifying individual units of the application, and identifying and quantifying relationships between the units. The mechanism by which the instrumentation package is attached to the application is described in detail below.
The illustrated ADPS requires knowledge of the structure and behavior of the target application. Data is gathered or supplied on how the application can be divided into units and how those units interact. ADPS functionality and effectiveness are limited by the granularity of distribution units, availability of structural metadata to identify units, choice of application analysis technique, representation of communication information, and mechanisms for determining location constraints on application units.
Granularity of Distributable Units
The granularity at which an application is divisible severely impacts the potential for improving performance of its distribution. Distribution granularity dictates the smallest independently distributable unit of the application. The number of potential distributions is inversely related to the distribution granularity. If the number of distributions is insufficient, none may offer good performance. However, if the granularity is too small, the tasks of choosing and realizing a distribution may become prohibitively expensive.
Perhaps even more importantly, the choice of partitioning unit shapes the relationships between partitioned granules. For instance, many distributed share memory (DSM) systems partition programs into VM pages. A single VM page often contains objects whose only commonality is their locality in creation time. The relationship between adjacent VM pages may be even more tenuous. Ideally, data within a distribution granule will exhibit good temporal and contextual locality.
The illustrated ADPS cannot choose granularity directly. The choice of distribution granularity is determined by the choice of operating environment. For instance, the distribution granularity in COIGN is a direct result of implementing the system on COM. An ideal environment for automatic distributed partitioning should provide a granularity of distribution with sufficient options to make automated partitioning worthwhile. The ideal granularity should match available metadata and provide a good "fit" to the application's structure.
Structural Metadata to Identify Units and Manage Communication
Distributed partitioning divides an application into units. Measurement of communication between units and division of units require access to appropriate metadata describing program structure. Program metadata can be derived from any of several sources including a compiler intermediate representation (IR), application debugging information, an interface definition language (IDL), and memory access data from the virtual memory (VM) system. Structural metadata provides the illustrated ADPS with sufficient information to separate application units and to manage code and data interactions among remote units of the application.
For example, in the COIGN system, IDL metadata and type libraries are provided by the Microsoft IDL compiler. IDL metadata is used to identify the number and type of arguments passed to and from interface functions. IDL metadata facilitates the identification and separation of components. Further, during distributed execution, IDL metadata is used to create proxies and stubs for cross-process and cross-machine communication.
Alternatively, other types of structural or program metadata can be used to identify application units.
Dynamic Application Analysis The illustrated ADPS generates the application description set 220. To do so, the illustrated ADPS can analyze (step 210) the structure of the application 200 and the communication between identified units of the application 200.
The choice of application analysis technique determines the type of application behavior visible to an ADPS. To work satisfactorily on applications in which application units are dynamically created and destroyed, a fully functional ADPS requires whole program analysis with complete information about the application's units, their dynamic instantiation relationships, and their communication patterns.
Dynamic analysis provides insight into an application's run-time behavior. The word "dynamic," as it is used here, refers to the use of run-time analysis as opposed to static analysis to gather data about the application. Major drawbacks of dynamic analysis are the difficulty of instrumenting an existing application and the potential perturbation of application execution by the instrumentation. Techniques such as a sampling or profiling reduce the cost of instrumentation. In sampling, from a limited set of application executions, a generalized model of application behavior is extrapolated. Sampling is only statistically accurate. In profiling, an application is executed in a series of expected situations. Profiling requires that profile scenarios accurately represent the day-to-day usage of the application. A scenario is a set of conditions and inputs under which an application is run. In the COIGN system, scenario-based profiling can be used to estimate an application's run-time behavior.
Referring to FIG. 7, scenario-based profiling of an application 200 to generate an application description set 220 is described. At step 202, structural metadata describing the application 200 is obtained. This structural metadata can be provided by an external source, or generated by the illustrated ADPS, as described in the preceding section. During later dynamic analysis, structural metadata can be used to determine how much data is between units of an application. For example, in the COIGN system, IDL metadata can be used to exactly identify function parameters, then measure the size of those parameters. With accurate interception and access to structural information, communication measurement is a straightforward process.
At step 204, the application 200 is executed in a scenario meant to model the expected use of the application 200. During execution, the application behaves normally while the numbers, sizes, and endpoints of all inter-unit messages are measured. At step 206, the user decides if profiling is finished. The application can be run through an arbitrary number of profiling scenarios. After profiling of the application is completed, the results from the scenario-based profiling are written (step 208) to the application description set 220. The application description set 220 can include structural description of the application as well as description of communication between units of the application.
Through scenario-based profiling, an ADPS can create a profile for each application unit instantiated during profiling runs of the application. The profile identifies and quantifies communication between the application unit and other units. The collection of profiles for all units in the application, together with the records of communications between units, can be included within the application description set 220 and used to decide where units should be placed in the network.
Network-independent Representation
An ADPS partitions an application to minimize its distributed communication costs. A correct distributed partitioning decision requires both realistic information about the network on which the application will be distributed, and accurate information about communications between units of an application.
In the illustrated ADPS, an appropriate inter-unit cost representation for an application is network-independent, but also incorporates realistic analysis of distribution tradeoffs prior to distribution. For example, referring to FIG. 6, an application description set 220 comprising a network-independent abstraction of inter-unit communication costs of an application can be combined with an environment description set 230 comprising basic statistics about a physical network to calculate concrete, network-dependent communication costs. While the environment description set 230 can be generated at the same time as the application description set, it can also be generated before or after. The environment description set 230 can be generated immediately before the application is to be distributed in a distributed computing environment, in this way describing the most recent state of the environment.
Network-independent representations of communication costs provide an application with a great degree of flexibility to adapt to future changes in network topology including changes in the relative costs of bandwidth, latency, and machine resources. In this way, a single application can be optimally bound to different networks, and a single application can be optimally bound and re-bound to a changing network. The ADPS preserves application flexibility by insulating the programmer from the final distributed partitioning decision. The programmer is responsible for exposing as many partitioning choices as possible by dividing the application into distributable units, but the ADPS is responsible for correctly distributing the application units for a given execution of the application based on the network environment. In essence, the ADPS allows late binding of an application to a particular network and its topology.
Late binding of an application across a specific network is facilitated by two mechanisms, described in detail below. First, compression of information about application communication reduces ADPS run-time overhead during profiling, and thereby enables more accurate and efficient summarization of network-independent communication costs. Second, quick estimation of the latency and bandwidth of a network allows the ADPS to delay partitioning until current estimates are needed. Combined, these techniques make it possible to delay binding of a distribution to a network until the latest possible moment, thus facilitating automatic adaptation to new networks.
In an alternative embodiment, estimates of latency and bandwidth are periodically taken during execution of a distributed application. If the new estimates deviate beyond a preset threshold from previous estimates, the application is re-partitioned and distributed using the new estimates. In another embodiment, inter-unit communication is measured during distributed execution. If the communication characteristics of the distributed application deviate beyond a preset threshold from the communication characteristics used to determine the current distribution scheme, the distributed application is re-partitioned and re-distributed
Alternatively, at a time when the characteristics of the distributed application deviate beyond a preset threshold, a notification can be given to the user. In response to the notification, the user can re-bind the application or ignore the notification.
Communication Representation
In the illustrated ADPS, during scenario-based profiling, communication between the application units is measured. Later, the illustrated ADPS partitions the application by comparing the inter-unit communication costs and network costs of alternative distributions. Because precise distributed partitioning analysis requires an accurate picture of the cost to distribute each unit of an application, the illustrated ADPS requires an accurate picture of the communication between units of an application.
During scenario-based profiling, the illustrated ADPS can measure the number and size of communications sent between any two application units. Pertinent features describing an inter-unit message are the source unit, the destination unit, and the amount of data sent from source to destination. For practical reasons, it is important to minimize perturbation of the application by the illustrated ADPS during scenario-based profiling. While the illustrated ADPS might ideally log all data about every message, doing so would most likely have a severe impact on application execution during profiling. Moreover, data about application communication needs to be preserved until the application is actually partitioned. If the size of the communication data is extremely large, preserving it can be prohibitively expensive. An inclusive log of all messages can be extremely large. It is conceivable that an application scenario could involve millions of messages.
Rather than store this information in a lengthy trace file, in the COIGN system, the number and size of inter-unit messages is selectively summarized. Various techniques can be used to compress application communication information.
The communication log can be compressed somewhat by storing messages with the same source and destination in a single collection. The source and destination need only be written once with subsequent records containing the size of the message only. However, the communication log might still be prohibitively large.
The communication log can be compressed even farther by noting that the important feature of the message in the partitioning decision is not the size of the message, but rather the communication cost of the message. The communication log for a source-to-destination pair could be compressed into a single number by summing the cost of all messages. However, to preserve generality it is desirable to separate the network dependent portion of the communication costs from the network independent portion.
The cost of sending a message consists of a latency factor, which is fixed for all messages, and a bandwidth factor, which is a function of the message size. The correlation of message size to bandwidth is nearly linear. Assuming that the bandwidth-cost function is in fact linear, instead of storing each message size, an alternative ADPS according to the invention stores the number of messages and the sum of the message sizes, as shown in the following equation 1: ##EQU1##
Unfortunately, the bandwidth-cost function is not strictly linear for most networks. Instead, the bandwidth-cost function is made up of discontinuous, near-linear ranges. The discontinuities occur when a message of size n+1 requires one more network packet than a message of size n. Not coincidentally, the discontinuities are a function of the network maximum transmission unit (MTU) and the network protocols. Compressing message sizes under the assumption that the bandwidth-cost function is strictly linear introduces an average error of 15% for a 10BaseT Ethernet. Similar errors are introduced for other networks.
An alternative approach to compress the log of messages is to compress each near-linear sub-range separately. For example, all messages from 0 to 1350 bytes could be linearly compressed into the number of messages and sum of message lengths. All messages from 1351 to 2744 bytes could also be linearly compressed. All messages above some large threshold value could be linearly compressed as MTU-induced discontinuities become less pronounced. MTU-induced non-linearities in the bandwidth-cost function are much more important for small messages than for large messages. As messages become larger, the amortized cost of each additional network packet becomes minimal. Unfortunately, compression based on the near-linear sub-ranges of a specific network is network dependent, which is something to be avoided.
Rather than linearly compress sub-ranges based on the MTU of a specific network, the ADPS of the present invention can linearly compress a number of exponentially larger sub-ranges starting with a very small range. For each sub-range, the decompression algorithm (i.e., the algorithm to calculate the cost of the compressed messages) is given by the following equation 2: ##EQU2##
where ##EQU3##
Latency.sub.small =Latency of the smallest message size in the sub-range,
Latency.sub.large =Latency of the largest message size in the sub-range,
Size.sub.small =Size of the smallest message in the sub-range, and
Size.sub.large =Size of the largest message in the sub-range.
In the COIGN system, the following sub-ranges for network-independent linear compression are used: 0-31 bytes, 32-63 bytes, 64-127 bytes, 128-255 bytes, 256-511 bytes, 512-1023 bytes, 1024-2047 bytes, 2048-4095 bytes, and 4096 bytes and larger. Compressing with these sub-ranges and then calculating values results in an average error of just over 1% for a 10BaseT Ethernet.
Determining Location Constraints
An ADPS can consider location constraints when partitioning application units for distribution. All prior work in ADPS systems has relied on programmer intervention to determine location constraints for application units. In the illustrated ADPS, location constraints can be desirably automatically detected and recorded, freeing the programmer from the task of identifying, tracking, and indicating location constraints.
Per-unit location constraints indicate which application units run better on a particular machine of the network or will not run at all if removed from a particular machine. The most common form of per-unit constraint is application unit communication through second-class communication mechanisms. A typical example of a second-class communication mechanism is a Unix file descriptor. The file descriptor represents a communication channel between the operating system and application. The file descriptor is a second-class mechanism because it cannot be directly distributed with first-class mechanisms, such as shared memory in a DSM system or interfaces in COM. The file descriptor implicitly constrains program location. In the COIGN system, system service libraries called by application units are analyzed to automatically detect second-class communication mechanisms and other per-unit location constraints. Alternatively, per-unit location constraints can be automatically detected by analyzing other application unit interactions with system resources.
Pair-wise location constraints indicate which combinations of application units must be located together. Pair-wise distribution constraints cannot be violated without breaking the application. For example, in COM, pair-wise constraints occur when two components must be co-located because they communicate either through an undocumented interface or through an interface that is not remotable because it uses opaque data types. In the COIGN system, pair-wise constraints are automatically detected during analysis of interaction between application units. If communication (e.g., function call parameters, data types) between two application units is not understood well enough to quantify the communication during profiling, a pair-wise location constraint is placed upon the two application units. Alternatively, if communication between the two application units is not understood well enough to remote the interaction (e.g., by marshalling and unmarshalling parameters over processes or machines) during distributed execution, a pair-wise location constraint is placed upon the two application units.
Decision: Deciding How the Application Should be Distributed
While an application can be partitioned in many ways, not all of them will yield equivalent performance. Application distributions that reduce the number and size of distributed messages are most likely to exhibit good performance. Because distributed communication is much more expensive than local communication, a distribution should minimize the amount of inter-machine communication. In addition to communication overhead, the illustrated ADPS can take into consideration relative computation costs and resource availability. A simple classification algorithm can be used to generate a distribution scheme 250 from an application description set 220 and an environment description set 230. Abstractly, the distribution decision consists of a communication model and cost metric that encode the decision problem for a particular application on a particular network, and an algorithm for optimizing the model.
An ADPS can model the tradeoffs between candidate distributions. Distribution costs can be modeled either directly or indirectly. Direct models specifically include communications costs between application units and resource availability. Indirect models consider contributing factors such as data or temporal locality. The choice of model determines which kinds of input data are required and which factors the optimizing algorithm maximizes. One very useful model of the distribution problem represents the application as a connected graph. Nodes represent units of the application and edges represent interactions between units. Edges are weighted with the relative cost of the interaction if remote.
Distribution Optimization Algorithms
The distribution optimization algorithm accepts a model of the decision problem and maps it onto a computer network. After all data has been gathered, it is the optimization algorithm that decides where application units will be placed in the network. In the COIGN system, the problem of deciding where to place application units is mapped to the common problem of cutting a commodity flow network. As described below with reference to FIG. 8, the application units and inter-unit communication form a commodity flow network. After this mapping, known graph-cutting algorithms can be used for automatic distributed partitioning.
A commodity flow is a directed graph 250 G=(N,E) with two special nodes (s 251 and t 252) designated respectively the source and sink. A steady supply of a commodity is produced by the source s 251, flows through the graph 250, and is consumed by the sink t 252. The graph 250 contains an arbitrary number of nodes 253 through which the commodity flows. Each node 253 may be connected to another node 253 by an edge 254. A node 253 may be connected to an arbitrary number of other nodes. Each edge 254 of the graph 250 has a capacity 255 that determines how much of the commodity may flow through it at a given time. The total flow through the graph is limited by the aggregate edge capacity 256. An important concept related to commodity flows is the cut 258. A cut (S, T) of a flow network G=(N,E) is a partition of the nodes N into two sets, S and T, such that the source s.epsilon.S and the sink t.epsilon.T and for all n.epsilon.N, n.epsilon.S or n.epsilon.T. The capacity of a cut 258 is the capacity of all of the edges connecting S to T; in other words, the capacity of the edges that cross the cut 258. A minimum cut is a cut of the commodity-flow graph with the smallest capacity.
In the case of a simple client-server network, the optimization algorithm can be a MIN-CUT MAX-FLOW algorithm, a type of optimization algorithm known in the art. The MIN-CUT MAX-FLOW theorem states that the capacity of the minimum cut is equal to the maximum flow through the flow graph. The capacity of the MIN-CUT is determined by the same edges that constrain the MAX-FLOW. The most efficient known algorithms to solve the MIN-CUT MAX-FLOW problem belong to the preflow-push family. The basic idea of the preflow-push algorithms is to use an iterative technique in which the commodity (limited by edge capacities) is pushed breadth-first through each edge from the source 251 to the sink 252. Excess commodity (when more commodity flows into a node than flows out) is iteratively pushed back to the sink again using a breadth-first algorithm. The simplest preflow-push algorithm runs in O(N.sup.2 E) time. Another algorithm used to partition client-server application across two machines, the lift-to-front algorithm, is a known preflow-push algorithm that runs in time O(N.sup.3), which is asymptotically at least as good as O(N.sup.2 E). The best known pre-flow push algorithm to date runs in time O(NE log (N.sup.2 /E)). Alternatively, other known optimization algorithms can be applied to a model of the decision problem.
While the problem of partitioning a graph into two sets (one containing the source and one containing the sink) can be solved in polynomial time, partitioning a graph into three or more sets (creating a multi-way cut) according to known algorithms in the general case is NP-hard. For this reason, practical multi-way graph cutting relies on approximation algorithms known in the art.
In the COIGN system, the algorithm to map a client-server distributed partitioning problem onto the MIN-CUT problem is as follows: Create one node for each unit in the application. Create one edge between every pair of communication units. The weight on the edge should be the difference between communication cost (communication time) for the remote case (when the two application units are placed on separate machines) and the local case (when the two application units are placed on the same machine). Create two additional nodes: the source and the sink. The source represents the client. For each application unit that must reside on the client-for instance, because it directly accesses GUI functions-create an edge with infinite weight from the source to the application unit. For each application unit that must reside on the server-because it directly accesses storage-create an edge with infinite weight between the sink and the application unit. Find the minimum cut of the graph. Since the minimum cut contains edges with the smallest weights (capacities), those edges represent the line of minimum communication between the client and server.
Each edge in the commodity-flow graph effectively represents the cost in time of distributing that edge. Because the common currency of graph edges is time, other time-based factors that affect distribution choice can be mapped readily onto the same MIN-CUT problem with communication costs. A good example is the problem of deciding where to place application units when client and server have different speed processors. For this case, two additional edges are attached to each application units. An edge from the application unit to the source s has a weight equal to the execution time of the application unit on the server. A second edge from the application unit to the sink has a weight equal to the execution time of the application unit on the client.
Each "computation" edge represents the cost in execution time if application unit is moved to the other computer. The MIN-CUT algorithm will cut through the edge that is least expensive (when considered with the other edges in the graph), thus leaving the application unit attached to the computer on which its aggregate communication and computation time is the lowest.
Each of the edges in the commodity flow graph is weighted with the same linear "currency". Because communication costs are most readily converted into time, the graph can be augmented with other time-based costs. In an ideal environment, one would also like to map discontinuous features into the graph problem. A common influencing factor in the choice of distribution is memory overhead. It is often desirable to keep memory footprint per client to a minimum on the server in order to maximize scalability of the server across multiple clients. Similarly, a client may not have enough memory to accommodate all application units that would ideally be placed upon it if considering time-based costs alone. The only known method to map memory overhead onto the graph-cutting problem uses a multi-commodity flow graph. Unfortunately, multi-commodity flow graphs are provable NP-complete in the general case.
Choosing a Distribution Online
In the illustrated ADPS, accurate values of latency and bandwidth for a particular network can be quickly estimated using a small number of samples, enabling adaptation to changes in network topology including changes in the relative costs of bandwidth, latency, and machine resources.
A correct distributed partitioning decision requires realistic information about the network on which the application will be distributed. If all distributed partitioning decisions are made offline, data for a particular network can be gathered from a large number of samples. For example, average latency and bandwidth values for a network can be derived from a large number of test packets sent on the network. In a dynamic environment where bandwidth and network availability can change from one execution to another, or within a given execution, it is desirable to make distributed partitioning decisions online at application startup. Data for online decision-making is gathered while the user waits. This creates a serious constraint on the number of samples used to determine available latency and bandwidth and model of network communication costs.
An ADPS minimizes communication costs between distributed application units by comparing alternative distributions. When comparing two application distributions, the communication costs in the first distribution are compared with the communication costs in the second distribution. The communication cost for any message is composed of two sub-costs: a fixed sub-cost due to network latency and a variable sub-cost due to network bandwidth. For some message m, the cost can be represented according to the following equation 3: ##EQU4##
The cost of an application distribution is the sum of the costs of all n messages sent between the partitioned application units given by the following equation 4: ##EQU5##
Measuring the real communication costs for a given network is extremely simple in theory, but somewhat error-prone in practice. For instance, to measure the average latency of a network, one sends a number of messages from one machine to another and back. One can compute the average round-trip time from either individual round trips using the following equation 5: ##EQU6##
or from the cumulative time for all of the round trips using the following equation 6: ##EQU7##
In practice, the round-trip time for a packet is unpredictable, making it hard to estimate average network behavior. This is particularly true for IP-based networks. Consider the round trip for a typical network message. The application initiates a message by creating a packet and invoking the operating system. The message passes through various layers in a protocol stack before the operating system eventually invokes the network interface. While travelling through the protocol stack, the message may be delayed by cache faults in the memory hierarchy. The network interface places the message onto the network medium. In many cases, such as shared medium token-ring or Ethernet, the network adapter may have to wait before actually transmitting the message. The message may travel over multiple physical networks; passing through routers to cross networks. At any router, the message may be dropped due to insufficient queue capacity on the router, forcing a re-transmission. When the message finally arrives at the receiver, it is placed in an incoming buffer. Again, the message may be dropped if the receiver has insufficient buffer capacity. In fact, the vast majority of message losses in typical networks are due to insufficient buffer capacity on the receiving machine. The network interface alerts the operating system, which picks up the message, passes it through the protocol stack, and finally delivers it to the receiving process. The receiving process takes appropriate action, then returns a reply to the sending process. The reply may wind its way back to the original process only to find that the original process was rescheduled after losing its scheduling quantum.
A message may be delayed at any point in the journey from the sender to the receiver and back. By measuring average round-trip time, an ADPS in fact measures the cumulative average effect of each source of delay. The more sources of spurious delay, the more measurements must be taken in order to calculate accurately the average round-trip time. Unfortunately, it takes time to make each network measurement. If network performance is unstable over time, then individual measurements will be unstable and the ADPS will therefore need more measurements to obtain an accurate view of current network performance. In contrast to average latency, minimum latency remains quite stable throughout all of the sources of delay typically introduced in networks. Stability in calculating the minimum network latency hints at the stochastic nature of packet-switched networks. No matter how heavy traffic is on a network, there are almost always a few packets that travel through the network at peak speeds. In fact, short-term performance of packet-switched networks is extremely unpredictable. If this were not the case, almost all packets would take a long time to travel through a heavily used network. In other words in a non-stochastic network, average latency and minimum latency would converge. Moreover, minimum latency fairly accurately tracks average latency for most networks.
In the illustrated ADPS, minimum latency and maximum bandwidth can be quickly measured with a short-term sample of measurements because even in congested networks, a few measurement packets pass through undelayed. Moreover, because minimum latency and maximum bandwidth reasonably track average values, minimum latency and maximum bandwidth values can be used in the illustrated ADPS.
Alternatively, an ADPS can utilize a combination of long-term values and short-term values. First, the ADPS can compute the average latency and bandwidth over an entire usage cycle--either a full day or a full week--and partition the application once accordingly. At the same time, the ADPS can create a library of stored average latency and bandwidth numbers--say one set of averages for each hour in the day--and depending on the time of day, partition the application according to the pre-computed network statistics. Second, after quickly estimating minimum latency and maximum bandwidth, these values can be matched to the closest stored average latency and bandwidth values, and the application then partitioned accordingly.
Distribution: Achieving a Chosen Distribution
Ultimately, an ADPS modifies the execution of the application to achieve a desired distribution. In the COIGN system, described in detail below, COIGN modifies the application by inserting an instrumentation package specially designed for distributing the application according to the desired distribution. This instrumentation package can be included with the instrumentation package used to identify units and measure communication, or can be a separate, lighter overhead package. Once the application is instrumented, achieving a distribution consists of two important steps: classifying application units and distributing them to the correct machine.
In general, through scenario-based profiling or static analysis, the illustrated ADPS creates a profile for each application unit instantiated. The profile classifies the application unit in order to characterize the application unit's communication with other units during profiling and any constraints on its location. Information from the profiling scenarios or static analysis is generalized to predict application behavior for later executions. A mapping of generalized application unit profiles to specific machines in the network is generated. Application units instantiated during application execution are then matched to similar application unit profiles, and located on the appropriate machine in the network. The actual distribution is an approximate solution to the distributed partitioning problem: the optimal solution for a particular application execution can only be determined after execution has completed. The underlying assumption of automatic distributed partitioning is that past profiles are statistically accurate in describing future application executions. If, in fact, past profiles accurately predict future application executions, then future executions can be partitioned using the distribution derived from the profiles.
Difficulties in classification by profile arise when application units are dynamic objects, such as COM components, for example. Component lifetimes are dynamic. A component may be instantiated or deleted at almost any point in program execution. Multiple instances of the same static type of component may exist concurrently. Moreover, separate instances of the same static type of component may have vastly different behavior and communication patterns due to their different usage contexts. For example, a single component in the document processing application, Octarine, is instantiated multiple times in a typical execution. Some instances hold references to operations invoked by menu commands. Some instances hold references to parts of a document including footers, headers, and body. Still other instances hold references to components in dialog boxes or spreadsheet cells. Two components with the same static type and similar communication patterns may need to be placed on separate machines if their sets of communicating partners are significantly different. In applications that are input-driven, user input typically drives the dynamic instantiation of application components. For this reason, component behavior varies tremendously between executions.
Component instances need to be classified not by their static type, but rather by their behavior and "where" they fit into the application. In essence, an instance needs to be classified by its usage context, e.g., the state of the application at the time the component is instantiated. The context in which a component is used is highly determinative of its pattern of communication with other components and the quantity of data communicated to other components.
Identification by Dynamic Classification
According to one aspect of the illustrated ADPS, in profiling and for distribution, application units are dynamically classified using the application state and the identities of the application units.
"Dynamic" classification refers to classification incorporating information about the dynamic state of the application. For example, a component is instantiated if a client uses it, so the client exists in some form, and information about the client can be used to classify the component. Parts of the dynamic state of the application can be approximated from dynamic structures which store application information, e.g., an execution call stack. A component can be dynamically classified at the time of instantiation using available contextual information such as the execution call stack, the arguments to the instantiation function, and the identities of previously instantiated components. An application's entire state (or at least an approximation thereof is available at the time of component instantiation to aid in classification. However, to be tractable, component classification uses only a limited subset of the application state. Contextual information can be summarized as a call chain including information about the sequence of usage of other components that preceded the component being classified.
The illustrated ADPS uses component identities to improve dynamic classification. A component is identified by a CLSID of the component, or by an identifier of the instance of the component. In the COIGN system, an instance identifier is supplied by instrumentation. Alternatively, an instance identifier can be provided for a component as part of an object model.
In the COIGN system, a component call chain (CCC) is used for dynamic classification. Entries in a CCC belong to a sparse, two-dimensional space: the product of the ca |
