Ordered message reception in a distributed data processing system6327630Abstract A complex computing system has a plurality of nodes interconnected by channels through which data messages are exchanged. The underlying principle is that after arrival at a node of a message, delivery of that message is delayed until after delivery and consequences of all more senior messages which affect the node. The messages are progressively timestamped at each node so that each time stamp contains generation by generation indicators of the origin of the associated message. The seniority of that message is uniquely determined thereby and total ordering of the messages can be achieved. When comparing timestamps for such ordering, comparison of respective generation indicators is necessary only until there is a distinction. Claims What is claimed is: Description This invention relates to complex computing systems. It was developed primarily to answer a problem with distributed systems, but it has been realised that it is equally applicable to systems which, are not normally considered to be distributed, such as a multi-processor computer. Although their physical separation may be negligible, nonetheless the processors are distinct and form a "distributed" system within the computer to which this invention is applicable.
class TTime ( friend TTime bool operator<( TTime
t0, TTime t1 ) {
public: if( t0.RealClock < t1.RealClock)
return TRUE;
TTime(): if( t1.RealClock < t0.RealClock )
return FALSE;
RealClock( 0.0 ), if( t0.Input < t1.Input ) return
TRUE;
Input( 0 ), if( t1.Input < t0.Input ) return
FALSE;
Length(0) {} for( unsigned i=0; i<min (
t0.Length, t1.Length ); i++ ) {
TTime( float realclock, unsigned input ): if(t0.Path[i] < t1.Path[i]
return TRUE;
RealClock( realclock ), if( t1.Path[i] <t0.Path[i] )
return FALSE;
Input( input ), }
Length(0) {} return t0.Length < t1.Length;
void AddNewProcess{} }
{Path[ Length++ ] = 0;}
void operator++() { Path[Length-1]++; }
friend bool operator<( TTime t0, TTime t1 );
private:
float RealClock;
unsigned Input;
unsigned Path[ MAX_PATH ];
unsigned Length;
The more processes handle a message, the longer its path grows. Potentially, if there is a cycle in the system, paths can become arbitrarily long. The implementation should use (when possible) dynamic allocation to avoid the arbitrary upper limit on path Each node or message handler invocation in the distributed system is structured as shown in FIG. 2. All functionality related to support for the time model resides in the time service, so that a process does not know anything about the time model. Each time it finishes handling a message it informs its time service (Done signal or event (e) above). The time service has a local clock T of type TTime which is updated whenever a message is received or sent by its process. Initially the local clock has a value given by a default constructor and it is kept and used even while the process is idle. The "timestamp assigner" at the border with the system's environment has a real clock synchronized with the clocks of all other timestamp assigners, and a unique input identifier. `Synchronized` here is understood to mean adequately synchronized, for example by means of the network time protocol as described in the Internet Engineering Task Force's Network Working Group's Request for Comments 1305 entitled `Network Time Protocol (Version 3) Specification, Implementation and Analysis` by David L Mills of the Univ. of Delaware published by the IETF in March 1992. Each time an external message enters the system it gets a unique timestamp constructed from these two values (see the second constructor of TTime). It is assumed that the real clock progresses between each two messages. Input messages are not delivered to the process until there are messages present on all inputs. Once this condition holds, the local clock is set to the timestamp of the most senior message, the new process is added to the path, and the most senior message is delivered. Every output message sent by the process while handling this input is timestamped by the time service with the current value of the local clock; and then the clock is incremented. The next message can be delivered only after the process explicitly notifies the time service that it has finished with the previous one, is idle and waiting (Done). The corresponding message sequence is shown in FIG. 3. The basic algorithm Alg. 1 of the time service is shown in the table below.
Initial slate: Idle.
Event Action
Slate Idle
Input message arrives .vertline..vertline.( there are messages
on all Inputs) // If all inputs are non-empty,
DeliverTheOldesIMessage(); // delivery is
possible
State Handling Message
Process sends output message Send it with the timestamp T;
T++; //increment the last time in the
timestamp
Done .vertline..vertline.( there are messages
on all Inputs) { // If all inputs are still non-empty,
DeliverTheOldesIMessage(): // deliver the
next oldest message
return;
}
Next state = Idle;
Functions
vold DeilverTheOldesIMessage() {
T = timestamp of the oldest message: // First, the local clock is set to
the value of the oldest
T.AddNewProcess(); // timestamp, and a new process is added
to the path in u
Deliver( the oldest message):
Next slate = Handling Message;
}
This algorithm ensures that input messages are delivered to each process in the order of their timestamps, and that output messages are sent by each process in the order of their timestamps. Thus, the time service as described above fully implements the time model. However, a distributed system containing a cycle will not work, as all time services in the cycle will always be missing at least one input message. Also, rare messages, either on an external input or on an internal process-to-process connection, may significantly slow down the whole system. Channel flushing can solve both these problems. Channel flushing is a mechanism for ensuring that a message can be accepted. The principle is to send auxiliary messages that enable the time service to prove that no message will arrive which is earlier than one awaiting delivery. Hence the waiting message can be delivered. There are two kinds, namely `sender channel flushing`, in which the sending end initiates channel flushing when the channel has been left unused for too long, and `receiver channel flushing`, in which the receiving end initiates channel flushing when it has an outstanding message that has been awaiting delivery for too long. Receiver channel flushing will be considered first, in conjunction with FIG. 4. For simplicity, timestamps and clocks are represented by single integers. Suppose for a certain period of time the two lower inputs of the process C are empty while there is a message with timestamp 23 waiting on the upper input. The time service of C wants to deliver the message as soon as possible, but it cannot do so until it proves that those messages that will eventually arrive on the empty inputs will have greater timestamps. To prove it, C sends a channel flushing request "May I accept a message of time 23?" to B and F, both of which have to forward this request deeply into the system, until either a positive or negative response can be given. In fact, to verify that C can accept the message with the timestamp 23 in FIG. 4, it is enough to ask only the processes shown, since all inputs to the diagram are at times later than 23. The algorithm described below is a straightforward implementation of channel flushing. All channels in the system (which actually connect the time services of the processes) are bi-directional, since, besides the normal uni-directional messages, the channel flushing messages are sent along them in the reverse direction. These messages and the structure of the time service are shown in FIG. 5. The general idea is that each time the time service discovers that there are input messages waiting while some inputs are empty, it sets a flush timer. On the timeout event it starts the channel flush. It sends flush requests to all empty inputs, creates a (local) request record with the list of these inputs, and then waits for responses. If positive responses come from all inputs, the oldest message is delivered. If a negative response comes on any input the flush is cleared and re-scheduled. Requests from other time services are handled in the following way. First, the time service tries to reply using its local information (its local clock and the timestamps of waiting messages). If it is unable to do so, it creates a (remote) request record and forwards the request to all empty inputs. If all responses are positive, so is the one to the remote requester. Otherwise, the response is No. The algorithm Alg. 2 is presented in a table below. Again, the time service has two states: Idle and Handling Message. While in the latter state, the time service is only serving process output messages, whereas in the Idle state it does all channel flushing work for both itself and other processes. <t, Path, Inputs>represents a request record. [] is an empty path. New, Delete and Find are operations over an array of request records that maintain the local state of the receiver channel flush algorithm. P is the identifier of this process. T is the local clock.
Initial slate: Idle, Flush timer not set no request records.
Event Action
State Idle
Input message for( all <I, Path, inputs> ): Path .noteq. []) //
First update remote requests.
with time 4 if( I < 4) // Input i is younger than the request's
time.
arrives at input i YesForRequest ( <t, Path, Inputs>, i );
else // Input is older than the request's time.
NoForRequest ( <1, Path, Inputs > );
if( there are messages on all Inputs ) ( // Then, if all
inputs are non-empty,
CancelLocalFlushing();
DeliverTheOldestMessage(); // delivery is possible.
return:
}
if( the new message is the oldest one ) { // if this
message becomes the
CancelLocalFlushing(); // oldest one, a new focal
Set flush timer; // flushing must be scheduled.
return:
}
if ( Find( < t, [], Inputs>)) // Otherwise, if
there is a local request waiting
YesForRequest ( < t, [], Inputs >, l ): // for
this input, that means Yes.
Flush timeout StartFlusing( timestamp of the oldest message, []);
// Empty return path indicates that the request is
local.
<Your Next Time?, t, if( P is among
[P.sub.o...P.sub.n).vertline..vertline.t<T) ( // If request has made a
cycle - assume
[P.sub.o...P.sub.n ] > Yes.
flush request Send to output to P.sub.n : <Yes, t,
[P.sub.o...P.sub.n ]>; // Or, if local clock is already
return; // ahead of t, definitely Yes.
}
if( there is a message older than t on some input) (
Send to P.sub.n : < No, t, [P.sub.o...P.sub.n ]>;
// Then assume No.
return;
}
// This process is not able to answer immediately and it
starts flushing.
StartFlushing(t, [P.sub.o...P.sub.n ]);
< Yes, t, [P.sub.o...P.sub.n ]> if( Find( <t, [P.sub.o...P.sub.n
], Inputs >))
+ve response on input YesForRequest ( <t, [P.sub.o...P.sub.n ], Inputs
>, i)
i
< No, t, [P.sub.o...P.sub.n ]> if( Find( <t, [P.sub.o...P.sub.n ],
Inputs > ))
negative response NoForRequest ( < t, [P.sub.o...P.sub.n ], inputs
> )
Slate Handling message
Process sends output Send it with the timestamp T;
message
Done if( there are messages on all Inputs ) ( // If all
inputs are still non-empty.
DeliverTheOldestMessage(); // deliver the next oldest
message.
return;
}
if( there is a non-empty Input) // Otherwise, if there
is a non-empty input,
Set flush timer; // a new local flushing must be
scheduled.
Next state = idle;
Functions
void DeliverTheOldestMessage() {
T = maximum(T, timestamp of the oldest message); // Before delivery, the
local clock is set
T.AddNewProcess(); // to the value of the oldest timestamp, and
Deliver( the oldest message); // a new process is added to the path in
this value.
Next state = Handling Message;
}
void CancelLocalFlushing() { // Cancelling local flushing activity includes
Cancel flush timer; // cancelling the flush timer (just in case it is set)
if( Find( <t, [], inputs > )} // and deletion of a local request
record (if any).
Delete( <t, [], Inputs > );
}
void StartFlushing( TTime I, TPath path){ // Flushing upon local or remote
request starts with
for(all empty inputs) // sending requests to all empty inputs
Send to input: < Your next time?, t, Path+P >, // (P is added to
the return path)
New( <t, Path, Set of empty inputs > ); // and creation of a new
request record.
}
void YesForRequest ( TRecord < t, Path, Inputs >, Tinput .vertline.)
{ // On positive response on input i
if( i g inputs ) // If the record has already received this
return; // information then it doesn't care.
Inputs = Inputs .backslash. i; // The input is removed from the Inputs set
of the request record.
if ( Inputs == .0.) { // If this set becomes empty, no more responses are
needed.
Delete( < I, Path, Inputs > );
if( Path == []) // and if it is a local request
DeliverTheOldestMessage(); // the oldest message is delivered.
// Successful channel flushing.
else // Otherwise it is a remote request.
Send to the last process in the Path: < Yes, t, Path >; // Yes is
sent to the next process
} // in the return path.
}
void NoForRequest( TRecord < t, Path, Inputs > ) { // Negative
response - acted on immediately.
Delete( < 1, Path, Inputs > ); // The corresponding record is
deleted.
if (Path == []) // If it is a local request
// Unsuccessful channel flushing
Set flush timer; // then restart the flush timer,
else // Otherwise, for a remote request.
Send to the last process in the Path: < No, t, Path >; // No is
sent to the next process
} // in the return path.
To illustrate the work of the algorithm, consider the example in FIG. 4 in conjunction with one possible channel flushing message sequence as shown in FIG. 6. "?" denotes here "Your Next Time?", "Y" stands for Yes, and "N" for No. Sets near the vertical axes represent the sets of processes from which responses are still wanted (Inputs in the above terminology). "ok" means that a process is sure it will not be sending anything older than the timestamp of the request, (i.e. 23). "loop" means that a process has found itself in the return path of the request. It is evident that the channel flushing procedure consists of two waves: a fanning out request wave and a fanning in response wave. The request wave turns back and becomes the response wave as soon as the information needed for the initial request is found. The "timestamp assigner" at the border with the environment treats the channel flushing request in the following way.
RealClock() returns the real clock reading; Input is the unique external
input identifier.
Event Action
External input message arrives Send it with the timestamp TTime(
RealClock(), Input):
< Your Next Time7, t, (P.sub.o...P.sub.n)> if( t < (TTime(
RealClock(), Input )) // lft is older than local time
flush request Send < Yes, t, (P.sub.o...P.sub.n)>; //
Then definitely Yes
else
Send < No, t, (P.sub.o...P.sub.n)>; //
Otherwise assume No
Sender channel flushing is conceptually simpler, and significantly more efficient than receiver channel flushing, although it does not provide a solution to the loop problem. In sender channel flushing, each output channel of each process has a timeout associated with it. This timeout is reset each time a message is sent down the channel. The timeout can be either a logical timeout (i.e. triggered by some incoming message with a sufficiently later timestamp) or a physical timeout. If the timeout expires before being reset then a sender channel flush is initiated down that channel. The channel flush consists of a `non-message` which is sent down the output channel. The receiver can use it to advance the time of the channel by allowing the time service to accept earlier messages waiting on other channels. When the non-message is the next message to be accepted, then the time service simply discards it. However, the non-message, by advancing the receiver's local clock, can cause logical timeouts on output channels of the receiver; hence causing a cascading sender channel flush. The timestamp assigners also participate in sender channel flush; they have to use a physical timeout. In general, using both sender and receiver channel flushes is recommended; preferably with some sender channel flushes piggy backed upon receiver channel flush response messages. To provide usable middleware implementing the time model it is necessary to relax some of the more restrictive assumptions about the system being built. Three special processes that need to be created and integrated with such middleware are now considered, as is a full treatment of loops in the dataflow. The Bus The bus is a process that allows multiple output ports from any number of processes to be connected to multiple input ports on other processes. The term `bus` is taken from Harrison (A Novel Approach to Event Correlation, Hewlett-Packard Laboratories Report No. HPL-94-68 Bristol UK 1994) and is intended to convey the multiple access feature of a hardware bus. The bus implements a multicast as a sequence of unicasts, its operation being shown in FIG. 7. The output channels are ordered, (shown in the diagram as 1 2, 3, 4). When a message is delivered by the time service to any of the input channels, the bus outputs an identical message on each of its output channels in order. The time service computes the timestamp for these in the normal way, as shown. A bus acts as a message sequencer, ensuring that all recipients of a series of multicasts receive the messages in the same order (as shown). The Delay In a non-distributed system it may be possible to set a timer, and then have an event handler that is invoked when the timer runs out. This alarm can be seen as a spontaneous event. Within the time model, it must be ensured that spontaneous events have a unique time. The simplest way of achieving this is to treat spontaneous events just like external events. A timestamp is allocated to them using a real clock and a unique input identifier. Moreover, a process can schedule itself a spontaneous event at some future time which again will get a timestamp with real part coming from the scheduled time. Having thus enabled the scheduling of future events the delay component can be created as schematically shown in FIG. 8. For each input message the delay generates an output message at some constant amount of time, .delta., later. The time of the generated message is given by the sum of .delta. and the first time in the timestamp (the real time part). The rest of the path part of the timestamp is ignored. The input identifier part of the timestamp is changed from the original, 1, to the input identifier of the delay, 1'. There are large efficiency gains from fully integrating delays with the receiver and sender channel flush algorithms. The responses to flush requests should take the length of the delay into account, as should any relaying of flush requests through the delay. The Plumber The plumber (the topology manager) is a special process that manages the creation and destruction of channels and processes. The plumber has two connections with every process in the system. The first is for the process to make topology change requests to the plumber; the second is for the plumber to notify the process of topology changes that affect it (i.e. new channels created or old channels deleted). The plumber can create and delete processes that have no channels attached. The plumber has a specific minimum delay between receiving a topology change request and doing it. This is the key to a feasible solution to topology changes within this time model. The reason that topology changes are difficult for (pessimistic implementations of) the proposed time model is that for a process to be able to accept a message it must know that it is the oldest message that will arrive. If the topology is unknown then all other processes within the application must be asked if they might send an older message. This is implausible. The plumber acts as the single point one needs to ask about topology changes. Moreover, the minimum delay between the request for a topology change and its realisation ensures that the plumber does not need to ask backward to all other processes. For large systems, or for fault tolerance, multiple plumbers are needed, and these can be arranged hierarchically or in a peer-to-peer fashion. As with the delay process the plumber needs to be integrated with the channel flush algorithms. Loops Loops generate issues for the time model and typical loops generate a need for many auxiliary messages. A loop without a delay can only permit the processing of a single message at any one time anywhere within the loop (it is said that the loop "locksteps"). Three solutions to these problems are examined. Removing Loops The traditional design models, client/server, master/slave, encourage a control driven view of a distributed system, which leads to loops. A more data driven view of a system, like the data flow diagrams encouraged in structure analysis, is typically less loopy. Moreover, where a first cut has loops, a more detailed analysis of a distributed system may show that these loops are spurious. The data-flows, rather than feeding into one another, feed from one submodule to another and then out. For example, in FIG. 9, there is an apparent loop between process A and process B, (a flow from A feeds into B which feeds back into A). But when one looks at the sub-processes A1, A2, B1, B2 there are, in fact, no loops, only flows. Co-locate the Processes in a Loop It there is a loop for which other solutions are not appropriate, it will be found that only one process within the loop can be operational at any one time. It will normally be better to have this, and put all the processes in the loop on the same processor. There will be no penalty in terms of loss of parallelism. This approach will minimise the cost of the auxiliary messages, because they will now be local messages. Break the Loop Using a Delay Informally, the problem with a loop is feedback. Feedback happens when an input message to a process strongly causes another message (the feedback) to arrive later at the same process. Under the strong causality axiom, feedback is strongly caused by the original messages, and hence comes before all subsequent messages. Hence any process in a loop must, after processing every message, first ascertain whether there is any feedback, before proceeding to deal with any other input. A delay process is a restricted relaxation of strong causality, since each input to the delay does not strongly cause the output, but rather schedules the output to happen later. Hence, if there is a delay within the loop, then a process can know that any feedback will not arrive until after the duration of the delay. Hence it can accept other messages arriving before the feedback. A Difficult Case The example in FIG. 10 presents specific problems of both semantics and implementation for feedback. In each of the four cases we see a message with data a arriving at a bus B and being multicast to processes A and C. A responds to the message a by outputting a message with data .beta. which is fed back into the bus, and hence multicast to A and C. When it arrives at A no further feedback is produced. If the bus sends to C before A (FIG. 10a) then no issues arise: the original message is multicast to both parties, and then the feedback happens and is multicast. If, on the other hand, the bus sends to A before C then he feedback happens before the original message is sent to C. The order in which C sees the feedback and the original message is reversed (FIG. 1b). This indicates that strong causality and feedback require a re-entrant processing similar to recursive function calls. Such re-entrant processing breaks the atomicity of invocations and also needs significantly more channel flushing messages than the non-re-entrant algorithm that has been presented. The simplest form algorithm Alg. 1 would incorrectly output the later message from B to C before the earlier message (FIG. 10c). Without re-entrant processing there is a conflict between strong causality and the sent after relation. The later version algorithm Alg. 2, refines the DeliverTheoldestMessage function to ensure that all incoming messages are delayed (with the minimum necessary delay) until after the previous output (FIG. 10d). This implementation obeys the time model, but (silently) prohibits non-delayed feedbacks. At the theoretical level this obviates the necessity for re-entrancy and prefers the sent after relation to strong causality. At the engineering level, this can be seen as a compromise between the ideal of the time model and channel flushing costs. A slight more exhaustive account of the above is given in the priority documents accompanying this Application.
|
Same subclass Same class Consider this |
||||||||||