Method and system for selecting data for migration in a hierarchic data storage system using frequency distribution tables5664187Abstract A hierarchic disk array data storage system has multiple storage disks that define a physical storage space and a RAID management system that maps the physical storage space into two virtual storage spaces. The RAID-level virtual storage space presents the physical storage space as mirror and parity RAID areas where the mirror RAID areas store data according to RAID Level 1 and the parity RAID areas store data according to RAID Level 5. The application-level virtual storage space presents the physical storage space as multiple virtual blocks which can be associated with mirror and parity RAID areas. The RAID management system migrates the virtual blocks between the mirror and parity RAID areas to optimize performance and reliability. To select virtual blocks for migration from mirror to parity RAID areas, a frequency distribution table is created to characterize the virtual blocks according to a predetermined distribution criteria. A virtual block is selected by segmenting the frequency distribution table to identify virtual blocks. The application-level virtual storage space is searched until a virtual block that satisfies the selection criterion is located. The located virtual block is then migrated to the parity RAID area. Claims What is claimed is: Description FIELD OF THE INVENTION
______________________________________
P.sub.0 = Segment 0 .sym. Segment 1 .sym. Segment 2
= Disk 0 .sym. Disk 1 .sym. Disk 2
P.sub.1 = Segment 3 .sym. Segment 4 .sym. Segment 5
= Disk 0 .sym. Disk 1 .sym. Disk 3
P.sub.2 = Segment 6 .sym. Segment 7 .sym. Segment 8
= Disk 0 .sym. Disk 2 .sym. Disk 3
P.sub.3 = Segment 9 .sym. Segment 10 .sym. Segment 11
= Disk 1 .sym. Disk 2 .sym. Disk 3
______________________________________
Parity redundancy allows regeneration of data which becomes unavailable on one of the disks. For example, if the data in segment 5 becomes unavailable, its contents can be ascertained from segments 3 and 4 and the parity data in segment P. Parity storage is less expensive than mirror storage, but is also less reliable and has a lower performance. The disk arrangement of FIG. 1 is provided for conceptual purposes. In practice, the disk array 11 would simply have a plurality of disks 12 which are capable of storing data according to mirror and parity redundancy. Among the available storage space provided by all disks 12, a portion of that storage space would be allocated for mirror redundancy and another portion would be allocated for parity redundancy. Preferably, disks 12 are configured to contain plural, equal sized storage regions (referenced as numeral 35 in FIG. 4), wherein individual regions have multiple segments. The regions are grouped together to form RAID areas in one virtual view of the storage space. Additionally, another (host-defined) view of storage space is presented to the user or host so that the RAID areas and data redundancy storing techniques are transparent to the user or host. These features are discussed below in more detail with reference to FIG. 4. Data storage system 10 manages the "migration" of data between mirror and parity storage schemes. The management of both types of redundancy is coordinated by RAID management system 16 (FIG. 1). RAID management system 16 manages the two different types of RAID areas in the disk array as a memory hierarchy with the mirror RAID areas acting similar to a cache for the parity RAID areas. RAID management system 16 shifts, organizes, and otherwise manages the data between the mirror and parity RAID areas in accordance with a defined performance protocol. The process of moving data between the mirror and parity RAID areas is referred to as "migration". Data storage system 10 tries to place the more performance-critical data in the mirror RAID areas since this affords the highest performance and reliability. The performance protocols implemented by RAID management system 16 includes one of two preferred migration policies. According to the first migration policy, known as "access frequency", the most frequently accessed data on the hierarchic disk array is maintained in the mirror RAID area 18. Less frequently accessed data is maintained in the parity RAID area 22. According to a second migration policy, known as "access recency", the most recently accessed data is maintained in the mirror RAID area 18 while the less recently accessed data is stored in parity RAID area 22. Other performance protocols may be employed. Ideally, such protocols are defined based upon the specific computer application and the needs of the user. Additionally, the RAID management system 16 automatically "tunes" the storage resources of a data storage system according to a function of two parameters: size of the physical storage capacity and size of the present amount of user data being stored in the data storage system. Initially, all data is stored in mirror RAID areas because this affords the highest performance and reliability. As more data is added to the data storage system, the data is migrated between mirror RAID areas and parity RAID areas to optimize performance and reliability. As the data storage system approaches full capacity, more and more data is migrated to parity RAID areas in an effort to meet all demands by the user while still providing reliability through redundancy. Accordingly, the data storage system of this invention affords maximum flexibility and adaptation. It does not require the user to select a specific storage regime, but instead can adapt to any demand placed on it by the user. FIG. 4 illustrates a memory mapping of the available storage space of data storage system 10 as multiple tiers of mapped virtual storage space. Each vertically elongated rectangle in the diagram represents a view of the physical storage space. In this diagram, physical storage space 34 is referenced by two virtual storage views 40 and 50. Physical storage space 34 is represented by four disks (such as disks 12 in FIG. 1) referenced by numerals 0, 1, 2, and 3. The four rectangles associated with the disks represent a view of the physical storage space wherein disks 1, 2, and 3 have approximately equal storage capacity, and disk 0 has slightly less storage capacity. Example storage capacities for such disks are 1-3 Gigabytes. The storage space 34 is partitioned into areas A0, A1, A2, etc. Individual areas contain numerous stripes, such as stripes 0-Q in area A0. Individual areas also contain numerous regions 35. Regions 35 preferably consist of a selected number of uniform sized segments on every storage disk so that the regions are equal in size across the entire disk array. An example size of one region 35 is one Megabyte. The storage space of the disks are mapped into a first, intermediate, RAID-level virtual view 40 of the physical storage space 34. This first virtual view is conceptually a set of RAID areas which can be mapped to a second application view that represents a contiguously addressable storage space. The physical configuration and RAID views of the storage space are hidden from the application view. The RAID area storage space 40 is the view of storage that identifies the mirror and parity storage space. For instance, a RAID area 42 may represent a mirror RAID area of M allocation blocks 43 while RAID area 44 may represent a parity RAID area of N allocation blocks 45. The allocation blocks 43 and 45 are preferably equal sized, with an example size being 64 Kilobytes. These RAID areas relate to corresponding physical areas A0, A1, A2, etc., on the physical storage space 34. As an example, sixteen 64K allocation blocks 43 or 45 at the RAID virtual view can be mapped onto a single 1M region 35. The mirror and parity RAID areas may or may not consume the entire storage space 34 of the disk array. Accordingly, during certain applications, there may be unused and undesignated storage space that does not correspond to a particular RAID area. However, such storage space can be converted into a mirror or parity RAID area. It is also noted that the RAID areas are shown as being mapped into contiguous areas on the disk array, where each region associated with a RAID area is located at the same physical address on each storage disk. The RAID areas may alternatively be mapped into non-contiguous areas on the disk array as well. The storage space available in the RAID areas is mapped into a second, front end, application-level virtual view 50 which is a view of storage as defined by and presented to the user or host application program. When viewed by the user or host application program, the application-level virtual view 50 can represent a single large storage capacity indicative of the available storage space on storage disks 12. Virtual storage space 50 presents a view of a linear set of equal sized storage virtual blocks 52 and 53, referenced individually as 0, 1, 2, . . . J-1, J, J+1, . . . , L-1, L, L+1, . . . , etc. Virtual blocks 52 and 53 are preferably the same size as the allocation blocks in RAID area 40, with an example size being 64 Kilobytes. The virtual block storage space 50 is represented by a table of references or pointers (as represented by arrows 54) to allocation blocks in the view presented by RAID areas 40. Virtual blocks 52 and 53 at the application virtual view 50 are therefore associated with allocation blocks 43 and 45 at the RAID virtual view 40 via the pointers maintained in the virtual block table. There are at least two types of RAID areas that can be referenced from the virtual block table: mirror and parity. The RAID management system 16 can dynamically alter the configuration of the RAID areas over the physical storage space. The number of RAID areas for each type may be increased or decreased depending upon the amount of user data being stored in the system and the size of the physical disk storage space. As a result, the mapping of the RAID areas in the RAID-level virtual view 40 onto the disks and the mapping of the front end virtual view 50 to RAID view 40 are generally in a state of change. The memory map store in NVRAMs 21a and 21b (FIG. 1) maintains the current mapping information used by RAID management system 16 to map the RAID areas onto the disks, as well as the information employed to map between the two virtual views. As the RAID management system dynamically alters the RAID level mappings, it also updates the mapping information in the memory map store to reflect the alterations. The migration operation of memory system 10 will now be described with reference to FIGS. 1 and 4. For purposes of continuing explanation, virtual blocks 53 of the application-level virtual storage space 50 reference associated allocation blocks 45 in parity RAID area 44 stored in area A1 of physical storage space 34. Such virtual blocks 53 are referred to as "parity virtual blocks" while the associated allocation blocks 45 are referred to as "parity allocation blocks". Similarly, virtual blocks 52 reference associated allocation blocks 43 in mirror RAID area 42 stored in area A0 of physical storage space 34. Such virtual blocks 52 are referred to herein as "mirror virtual blocks" while the associated allocation blocks 43 are referred to as "mirror allocation blocks". In general, to migrate data from one RAID area to another, a virtual block associated with an allocation block of a first RAID level type (such as mirror or Level 1) is selected. Then, an unused allocation block representing a second RAID level type (such as parity or Level 5) is located. If an unused allocation block cannot be located, one is created. Data is next transferred from the allocation block previously associated with the selected virtual block to the unused allocation block which causes the data to undergo a redundancy level change. For example, data once stored according to mirror redundancy would now be stored according to parity redundancy, or vice versa. As a final step, the mapping 54 of the application-level virtual storage space 50 to the RAID-level virtual storage space 40 is modified and updated to reflect the shift of data. The selected virtual block that was formerly associated with an allocation block of the first RAID level type now references via an updated pointer an allocation block of the second RAID level type which contains the migrated data. Any mapping change occurring during this transaction would be updated in memory map store 21. The continuing discussion provides a more detailed explanation of migrating data between mirror and parity storage areas according to preferred methods and sequences of this invention. To migrate data from parity to mirror storage, the following sequence is employed: 1. The RAID management system locates an unused mirror allocation block 43 in a mirror RAID area 42. 2. If none can be found, the RAID management system creates a mirror allocation block (discussed below). 3. The RAID management system suspends new storage requests to the virtual block to be migrated. 4. The RAID management system waits until all active data storage requests to the virtual block are completed. 5. The data from the parity allocation block 45 associated with virtual block 53 is read into a temporary memory buffer. 6. The data is then written to the mirror allocation block 43 chosen in step 2. 7. The virtual block table is modified to reference the new location of the data in the mirror allocation block 43. 8. The suspended storage requests are resumed. According to the above procedure, a virtual block 53 in the application-level virtual view 50 migrated from parity to mirror storage. Relative to the intermediate virtual view 40, data has migrated from a parity allocation block 45 in parity RAID area 44 to a mirror allocation block 43 in mirror RAID area 42. In the physical storage space data has moved from area A1 to area A0. If an unused mirror allocation block cannot be located (step 1 above), the RAID management system tries the following preferred sequence of three techniques. First, the RAID management system will try to locate an unused (and thus undesignated) RAID area, which can be converted to a mirror RAID area without violating the system threshold of unused RAID-level storage that is needed to guarantee that migration can always proceed. If this fails and the system has more than the reserved amount of unused RAID-level storage, the system migrates data within parity storage to collect unused parity allocation blocks into unused RAID areas. If this migration yields an unused RAID area that can be converted to a mirror RAID area as above, then the system converts it to a mirror RAID area. Otherwise, the system alternately migrates data from mirror to parity storage, packs mirror storage, and converts unused RAID-level storage to parity until the system increases unused RAID-level storage sufficiently for the location of an unused mirror allocation block or a conversion of an unused RAID area to a mirror RAID area. Since mirror allocation blocks occupy more physical storage space than parity allocation blocks, this last technique will result in a net increase in the amount of unused RAID-level storage. The creation/conversion protocol used to locate and establish unused mirror allocation blocks is advantageous because it permits the RAID management system to selectively adjust the memory/allocation between parity and mirror areas according to the amount of user data and the size of physical storage space. As data usage and storage capacity vary, the RAID management system employs one or more of the above three techniques to maximize the amount of data held in mirror storage. The RAID management system attempts to avoid the situation in which a storage request must wait for the space-making sequence to yield an unused mirror allocation block by creating unused RAID areas during idle time. However, in some situations, storage requests may be suspended during the space-making sequence. The RAID management system configures the virtual block storage space in such a way that the virtual space will be smaller than the RAID view. This ensures that a free space equal to at least one RAID area is set aside for migration or other purposes. In this manner, the sequence of techniques will always yield an unused mirror allocation block. To migrate data from mirror to parity storage, the following sequence is employed: 1. The RAID management system chooses a virtual block from 52 to migrate from mirror to parity storage according to a migration policy such as access recency or access frequency. 2. The RAID management system locates an unused parity allocation block 45 in a parity RAID area 44. 3. If such a block cannot be found, space reserved for migration is converted to a parity RAID area according to the above described creation techniques. 4. New storage requests to the virtual block to be migrated are suspended. 5. The RAID management system waits until all active storage requests to the virtual block are completed. 6. Data is read from the mirror allocation block 43 associated with virtual block 52 into a temporary memory buffer. 7. The data is written to the chosen parity allocation block 45. 8. The virtual block table is modified to reference the new location of the data in parity allocation block 45. 9. Data requests to the virtual block are resumed. The above two nine-step sequences provide examples of how the memory system of this invention can operate to migrate data between two different levels of redundancy. Data is migrated between parity and mirror RAID areas using one of the two preferred migration policies of data access recency and data access frequency. In general operation, the data storage system of this invention wants to migrate the oldest or least frequently accessed virtual blocks from mirror to parity storage. Contrarily, it might be useful to migrate the newest or most frequently accessed virtual blocks from parity to mirror storage. One of the tasks of the data storage system is to select the appropriate virtual blocks for migration. According to this invention, the virtual blocks are tagged with a monotonically increasing mark to assist in the selection process. If the migration policy is access recency, the virtual blocks are tagged with a time stamp indicating when the data was stored therein. The time stamp is preferably generated by a counter in the disk array controller 14 which outputs an incremental tick about every minute (in one implementation, every 65.5 seconds). The counter thereby forms suitable means for marking such storage blocks, although other marking means may alternatively be employed. For access frequency, the virtual blocks are tagged with a frequency stamp indicating the number of times the data therein has been accessed. The frequency stamp is an integer value that is increased by one with each access of the corresponding virtual block. One straightforward technique for locating the appropriate virtual blocks for data migration is to sort the entire second virtual storage space 50 by time or frequency stamp. The sorting would yield the oldest/youngest or least frequently/most frequently accessed virtual blocks for migration between parity and mirror RAID areas. This technique is less preferred, however, because it takes considerable memory resource and processing time to sort all of the virtual blocks. The shortcomings of the sorting technique become more pronounced as the size of user data increases. The processing time is a non-linear function of the number of virtual blocks to be sorted. As the number of blocks increases, the processing time becomes longer. A second and more preferred technique of this invention is to statistically select suitable virtual blocks without resorting to a full sort. According to this technique, one or more frequency distribution tables are established by the RAID management system 16 to statistically characterize the virtual blocks in terms of their time or frequency stamps. The frequency distribution table(s) can then be processed to identify a certain stamp threshold that segments the older/younger or less frequently/more frequently accessed virtual blocks from the remaining population of virtual blocks. Once a segment stamp threshold is identified, the application-level virtual storage space can be quickly searched until a sufficient number of virtual blocks within the segment have been found. The search effort is then halted. The data in the located virtual blocks is manipulated to carry out the desired function, such as migration to other allocation blocks. The statistical selection technique optimizes both memory usage and processing time and is therefore preferred over the full sorting technique. FIGS. 5-12 demonstrate the various uses, creation, and maintenance of frequency distribution tables in hierarchic data storage systems according to this invention. These aspects and corresponding novel methods will be explained with reference to the flow diagram of FIGS. 10-12. Preferably, the methods described herein are implemented as firmware in RAID management system 16. FIG. 10 illustrates an initialization process. The first two steps 100 and 102 establish the RAID environment for the hierarchic data storage system. At step 100, RAID management system 16 maps the physical storage space provided by disk array 12 into RAID-level virtual storage space 40. The RAID-level virtual storage space presents the physical storage space as multiple RAID areas that store data according to various RAID levels, such as RAID Level 1 and RAID Level 5. At step 102, the RAID management system maps the RAID-level virtual storage space into application-level virtual storage space which presents the physical storage space as multiple virtual blocks. The virtual blocks are associated with corresponding RAID areas. Data is stored in the virtual blocks according to the RAID level of the associated RAID area. At step 104, the RAID management system creates one or more frequency distribution tables. Preferably, one frequency distribution table is created for each type of RAID area. The frequency distribution tables are preferably formed in a volatile RAM in the disk array controller 14 and rebuilt after each system initialization. Alternatively, the distribution table can be resident in the NVRAMs of memory map store 21 in disk array controller 14. The NVRAM implementation is preferred if it is foreseen that rebuilding the table after each system initialization is too time consuming. Aspects of these tables are described in more detail with reference to FIGS. 5 and 6. FIG. 5 illustrates a frequency distribution table 60 in a bar graph format and FIG. 6 illustrates the same table in a memory format conceptualization. Frequency distribution table 60 statistically characterizes the data stored on the data storage system according to a predetermined distribution criteria. Two preferred distribution criteria are data access recency or data access frequency. The table makes the characterization at the virtual block level in second virtual storage space 50. Thus, the table contains statistical information about the virtual blocks being used to store user data. Frequency distribution table 60 is preferably bounded with a fixed number of intervals. An example number of intervals is ten, as shown in FIGS. 5 and 6. The intervals are labeled generically as 1-10, but such intervals actually consist of a range of values from a beginning boundary value to an ending boundary value. If the distribution criteria is access recency, the intervals are temporal in scope where each interval covers a range of time stamps associated with the virtual blocks. An example interval consists of 1024 time stamps which, at 65.5 seconds per stamp, covers all of the virtual blocks used in a time period of approximately one day. If the distribution criteria is access frequency, the intervals relate to frequency values where each interval covers a range of frequency stamps associated with the virtual blocks. The various intervals and counts under the access frequency regime depend upon how actively the user accesses its data. The table contains the number of virtual blocks that occur within each interval. This is represented by the vertical bars and associated counts in FIG. 5. For example, thirty-two virtual blocks have a time or frequency stamp that falls within interval 1; forty-eight Virtual blocks have a time or frequency stamp that falls within interval 2; and so on. The table also tabulates a cumulative frequency or total number of virtual blocks to define the entire population. In this case, the total population consists of 566 virtual blocks. In FIG. 6, frequency distribution table 60 contains a first field 62 which stores the total population of 566. The table also contains a population queue 64, a free queue 66, and ten interval fields 68. Population queue 64 maintains the interval elements for the current population whereas free queue 66 tracks the unused interval elements. Each interval field 68 contains the count for the associated interval (i.e., thirty-two for interval 1, forty-eight for interval 2, and so on), the beginning boundary value, and the size of the interval. From the beginning boundary value and the interval size, the ending boundary value can be readily computed. The frequency distribution table is created during system initialization by scanning the virtual blocks contained in the application-level virtual storage space. In the beginning, population queue 64 is empty and free queue 66 contains ten unused interval elements. The count and beginning boundary value for all ten interval fields 68 default to zero. The size for each interval field is initialized to a default size, such as 1024. The table is then statistically constructed based upon the stamps, such as the time or frequency stamps, of the virtual blocks (step 106 in FIG. 9). The number of virtual blocks satisfying each interval is counted and input to the appropriate interval field 68. The stamp for each virtual block is sequentially checked (as indicated by the flow arrows in FIG. 6) against the interval fields 68 until a match occurs (step 107 in FIG. 9). For example, assume that the virtual block had a stamp that fell within the interval in the third interval field. The virtual block stamp would first be evaluated against the interval in the first interval field. Since the virtual block falls outside this interval field, it is next compared to the second interval field. It is noted that each interval field contains pointers to direct analysis to the address for the next interval field. The process continues field by field until the virtual block stamp is compared and found to fit in the appropriate interval field (which in this case, is the third interval field), whereupon the count for that interval field in incremented (step 108 in FIG. 9). It is noted that unused virtual blocks have no stamp associated with them, and thus are not represented in the table. By maintaining top and last interval pointer values, the population queue 64 can readily find and point to the top and last interval fields 68. As virtual blocks are scanned and the statistical information extracted, population queue 64 begins to fill and free queue 66 begins to empty. When the table contains a count in each interval, the free queue 66 will be empty (as denoted by the null or ground symbols in the free queue). It is noted that when two or more tables are desired (as described below in more detail), they can be created simultaneously during the same scanning step. Once the frequency distribution tables are established, they are ready for use. The flow diagrams of FIGS. 11 and 12 demonstrate two preferred modes of operation involving the frequency distribution tables, although other operations using the frequency distribution tables are likely. The first or host access mode of operation is to update the tables as the user accesses existing or new virtual blocks. The second or migration mode of operation concerns segmenting the tables to identify virtual blocks for use in migration between RAID areas and updating the frequency distribution tables. FIG. 11 shows the host access mode of operation wherein the tables are continually updated as existing and new virtual blocks are used to store user data. At step 109, it is first determined whether the user is accessing an existing virtual block or is creating a new virtual block. If the user is accessing an existing virtual block (i.e., the "no" branch from step 109), a new stamp is applied to the existing virtual block. The frequency distribution tables are updated to reflect the new stamp. This is done by decrementing the count of the interval containing the old stamp (step 110) and then incrementing the count of the interval containing the new stamp following step 111. FIG. 7 illustrates the process of updating a table when an existing virtual block is accessed. Here, a virtual block having a time stamp in the range of interval 6 is accessed. A new, most recent time stamp is then tagged to the virtual block. One significant advantage to having a monotonically increasing stamp is that the system can easily locate and handle the most recently/frequently accessed and least recently/frequently accessed virtual blocks because these blocks are located in the first and last intervals. The most recent time stamp therefore expectedly falls within the range of interval 10. The frequency distribution table 60 is updated by decrementing the count of interval 6 by one (80 to 79) and incrementing the count of interval 10 by one (76 to 77). With reference to FIG. 11, if the virtual block is new (i.e, the "yes" branch of step 109), analysis continues at decision step 111 in FIG. 11 to determine whether the new virtual block fits an existing interval within a table. Given the monotonic nature of stamping virtual blocks, the virtual block is simply compared to the last interval of the table. If the new virtual block fits the last interval (i.e., the "yes" path from decisional step 111), the count associated with that last interval is incremented (step 112). For instance, if a virtual block has a time stamp falling in the range of interval 10, the count of interval 10 is incremented from 76 to 77. On the other hand, if the virtual block has a stamp that does not satisfy any of the last intervals (i.e., the "no" path from decisional step 111 ), the RAID management system quickly checks if any intervals are free (step 113). This situation occurs when less than the set number of intervals (e.g., ten intervals in the continuing example) have been used to describe the frequency distribution of the last virtual blocks. If an interval is free (i.e., the "yes" branch from step 113), the new last interval is defined at step 116 to contain the new virtual block. When no free intervals exist (i.e., the "no" branch from step 113), a new interval must be created to accommodate the increasing distribution range. Creating a new interval requires some adjustment to the frequency distribution table because it is bounded with a fixed number of intervals. The RAID management system does not merely add a new interval. Instead, the RAID management system first makes room within the table to accommodate the new interval. FIG. 8 illustrates the process of generating a new interval. The RAID management system first combines adjacent intervals to form a new joint interval (step 114 in FIG. 11). The new joint interval has a count that equals the sum of the counts of the two intervals that are joined. Preferably, the adjacent two intervals with the smallest combined count are selected. In FIG. 8, intervals 2 and 3 are joined to form a new interval 2 in table 60'. The count of the new interval 2 is 76 (i.e., 48+28=76). This process frees up space in the frequency distribution table for definition of a new last interval (step 116 in FIG. 11 ), such as new last interval 10. The count for the new last interval is then incremented to reflect the new virtual block (step 118). In table 60', the new last interval 10 has a count of one to signify statistically that the new virtual block fits that interval and did not fit the previous intervals. Through the joining process, the frequency distribution table remains bounded and easy to handle, and the statistical information therein is accurately maintained. As noted above, in the preferred implementation of this invention, one frequency distribution table is constructed for each RAID level. When the hierarchic disk array uses RAID Levels 1 and 5, a first frequency distribution table is created for virtual blocks associated with mirror RAID areas of RAID Level 1 and a second frequency distribution table is created for virtual blocks associated with parity RAID areas of RAID Level 5. If three or more RAID Levels are used, then three or more corresponding tables are created. An example use of this memory arrangement might be to migrate "middle-aged" virtual blocks to a level of storage that has performance characteristics most suitable for their particular data. FIG. 9 shows two frequency distribution tables 60 and 70. Upper table 60 is used to characterize virtual blocks that are associated with RAID Level 1 storage. Lower table 70 is used to characterize virtual blocks that are associated with RAID Level 5 storage. FIG. 9 also illustrates the effect of migration on the frequency distribution tables for the different RAID levels. The method for using the table for migration is shown in the migration mode of operation in FIG. 12. For purposes of this example, the tables in FIG. 9 are based upon the distribution criteria of data access recency and the intervals are temporal in scope. Assume that the RAID management system 16 wants to migrate "older" virtual block from mirror to parity storage to free up mirror storage for new data. In this case, the mirror RAID areas operate as the source RAID areas for data migration and the parity RAID areas operate as the target RAID areas (step 120 in FIG. 12). The RAID management system first divides the population of virtual blocks in table 60 into a plurality of equal size segments having approximately the same number of elements (step 122). For example, the RAID management system might divide the population into five segments, as shown at the top of table 60 in FIG. 9. For a population of 566, each segment should theoretically represent approximately 113 virtual blocks. The first segment spans intervals 1-4 because intervals 1-3 only have 105 virtual blocks, below the threshold 113 virtual blocks for each segment. The RAID management system then chooses the first segment since it will contain the older virtual blocks (step 124 in FIG. 12). Any virtual block having a time stamp falling within the temporal intervals referenced generically as 1-4 will satisfy this criterion. The RAID management system determines the uppermost boundary value of the first segment which, in this example, is the ending boundary time value of interval 4. The ending boundary time value is computed by taking the beginning boundary time value of interval 4 and adding the size of interval 4. The ending boundary time value is used as a threshold for comparison. The RAID management system then searches through the virtual blocks in the application-level virtual storage space to actually locate one or more virtual blocks (step 126). The time stamp of each virtual block is compared to the threshold time value to see if the virtual block has a time stamp earlier than the threshold time value, thereby placing the virtual block in the first segment. Once a sufficient number of virtual blocks are found, the search is halted. The located virtual blocks are then migrated from mirror storage to parity storage (step 128). The two frequency distribution tables 60 and 70 are updated to reflect the migration (step 130). Suppose that the migration of three virtual blocks previously represented in interval 2 of RAID Level 1 table 60 is now reflected in interval 10 of RAID Level 5 table 70. The count associated with interval 2 in table 60 is simply decremented by three (from 48 to 45) to reflect the loss of the virtual blocks. Similarly, the count associated with interval 10 in table 70 is incremented by three (from 47 to 50) to reflect the gain of the virtual blocks. It is noted that the time stamps of the three virtual blocks do not change during migration. Rather, only the statistical information for each RAID level is modified to reflect the change in distribution resulting from the data migration. It is noted that the migration is preferably performed in a continuous fashion, as opposed to a batch procedure. Accordingly, once a criterion-satisfying block is found, it is migrated, and the tables updated, in a block-by-block manner. The use of frequency distribution tables affords more advantages than the full sorting technique. The statistical approach yields significantly better performance than sorting methods, while employing significantly less memory resource. The process of segmenting the population, choosing a segment to identify virtual blocks, and then quickly searching to find the virtual blocks approximates the full sorting technique. Only it accomplishes the task in a more efficient and cost effective manner because it enables intelligent selection without expense. Additionally, the process can be stopped as soon as a sufficient number of virtual blocks are found in the identified segment. By stopping earlier, the truncated time frame is generally significantly less than that required to perform a full sort. Another benefit is that the processing time is a linear function of the number of virtual blocks (for searching) unlike the non-linear function of full sorting techniques. Thus, older/younger and less recently/more recently accessed virtual blocks can be more quickly identified with respect to the amount of user data stored on the disk array. The use of statistical frequency distribution tables can also be used to select data for movement between different types of storage devices. For example, if the data storage system was further equipped with a back-up storage mechanism, such as reel-to-reel tape storage, the statistical frequency distribution table could be used to select "ancient" storage blocks for data movement from the disk array to the back-up storage mechanism. In compliance with the statute, the invention has been described in language more or less specific as to structural and methodical features. It is to be understood, however, that the invention is not limited to the specific features shown and described, since the means herein disclosed comprise preferred forms of putting the invention into effect. The invention is, therefore, claimed in any of its forms or modifications within the proper scope of the appended claims appropriately interpreted in accordance with the doctrine of equivalents.
|
Same subclass Same class Consider this |
||||||||||