The present invention generally relates to the design of multi-level cache architectures in computer systems, and more specifically to the design of such architectures in which cache lines residing in a lower level cache may bypass an intermediate cache in the cache hierarchy when the cache lines are requested by and stored in a higher level cache.
One may begin by considering (for example) a computer system with a processor together with L1 (level 1) and L2 (level 2) caches. In the case of an L1 miss, it is clearly advantageous to bypass the L1 and load data from the L2 directly into processor registers in the case that the cache line, if it had been loaded into the L1, would never have been re-used before being replaced. In this example, the processor registers can be thought of as the highest level of the cache hierarchy, that is, as an L0 (level 0) cache. The advantages of bypassing a cache in certain cases have been recognized previously, and a number of techniques proposed for selective cache bypass, as described in more detail below. However, none of this work describes the use of predictive mechanisms at cache line granularity in order to selectively enable and disable cache bypass based on the recent usage history of cache lines for all cache lines residing in the L1 and L2 at any point in time.
In a conventional technique known as lazy cache promotion, two L1 misses for a given line are required before the line is loaded into the L1 (see Efficient Selective Caching through Lazy Cache Promotion, published electronically at IP.com, document ID IPCOM000008849D, Jul. 17, 2002). It is clear that in certain cases this method will lead to decreased system performance, since two misses are required to load a line into the L1 that would more advantageously have been loaded on the first reference; furthermore, usage information is not maintained (that is, when a given line is replaced in the L1, the information that it was previously loaded due to two subsequent misses is discarded).
Compiler-based techniques have also been proposed, that is, compile-time analysis is used to attempt to determine certain variables that map to cache lines which should bypass a cache in the cache hierarchy. Examples include (1) Compiler Managed Micro-cache Bypassing for High Performance EPIC Processors, Wu et al, pages 134-145, Proceedings of the 35th ACM/IEEE International Symposium on Microarchitecture, November 2002, and (2) Unified Management of Registers and Cache Using Liveness and Cache Bypass, Chi and Dietz, pages 344-355, Proceedings of the ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI), Portland, Oreg., Jun. 21-23, 1989. Possible disadvantages of these approaches include the coupling of compilers and compile-time analysis with details of the system's cache architecture, together with the maintenance and use at run-time by the cache controllers of compiler-generated information associated with particular programs with resulting complexity, together with identification of only a limited set of cache lines for which cache bypass may be advantageous.
Probabilistic methods have also been proposed for use in selective caching. In the publication Probabilistic Cache Replacement, Technical Report TR-13-95, Harvard University, 1995, a method is proposed in which caching of data from main memory is determined by a probabilistic boolean-valued function, where repeated misses to a word in memory increase its probability of being cached. One of the disadvantages of this approach is that multiple misses are required to increase the probability that a given memory word will be cached, which more advantageously could have been cached on the first miss.
Other approaches may be classified as instruction-based methods. Examples include (1) Managing Data Caches using Selective Cache Line Replacement, Tyson et al, International Journal of Parallel Programming, Vol. 25 No. 3, 1997, and (2) Automatic Cache Bypass for Instructions Exhibiting Poor Cache Hit Ratio, Jamshed H. Mirza, U.S. Pat. No. 5,625,793, Apr. 29, 1997. In these approaches, individual instructions may be identified (using offline analysis) and marked for which a performance advantage may be realized by not caching (in the L1) the target of such marked instructions. The possible disadvantages of these approaches are similar to those of compiler-based techniques.
Other related work includes methods in which a data reference pattern is detected. First, in the publication Reducing Conflicts in Direct-Mapped Caches with a Temporality-Based Design, Rivers and Davidson, pages 154-163, Proceedings of the 1996 International Conference on Parallel Processing, August 1996, a cache architecture design is described in which the L1 cache is partitioned into a direct mapped cache and a fully associative cache (used as a buffer); cache lines that are loaded into the direct mapped partition but not re-used before being replaced are tagged and then on subsequent references loaded into the fully associative buffer partition. Some possible disadvantages of this approach are that conflict misses for a given line in the direct mapped partition may be transient (that is, a line may typically be re-used but occasionally not re-used due to a non-recurring conflict), extra complexity associated with the fully associative buffer, possible over-utilization of the fully associative buffer in the case that an excessive number of lines are tagged, and lack of a mechanism for clearing the previously described tags.
Finally, in some cases sequential access by a program to a large array (for example) may cause a cache in the cache hierarchy to be loaded with large amounts of data that are not re-used. If such a pattern can be recognized, the data can bypass the cache. Such a technique is described in Data Processing with First Level Cache Bypassing After a Data Transfer Becomes Excessively Long, Thomas Leyrer, U.S. Pat. No. 5,729,713, Mar. 17, 1998. A primary disadvantage of this approach is that it is equally possible that the sequentially accessed data could be re-used multiple times; for a simple example, consider a program for matrix multiplication of two N×N matrices in which each column and row of the two respective matrices will be accessed N times (and in which the cache is large enough to hold the relevant data).
Accordingly, a need has been recognized in connection with overcoming the shortcomings and disadvantages as described hereinabove with respect to conventional arrangements, and in selectively enabling and disabling cache bypass at cache line granularity for a cache in a multi-level cache hierarchy using predictive mechanisms.
In accordance with at least one presently preferred embodiment of the present invention, in utilizing recent usage history, predictive mechanisms preferably selectively enable and disable cache bypass at cache line granularity, for all cache lines resident in the cache which may be bypassed and in the next lower-level cache.
In a computer system with a multi-level cache hierarchy, as described above it is desirable to have a cache bypass prediction mechanism that can predict when requested data should bypass a particular cache in the cache hierarchy, in order to improve overall system performance. Consider a computer system that includes level i−1, level i, and level i+1 caches C[i−1], C[i], and C[i+1], respectively. In the case that i=1, the cache C[0] will be considered to include the processor registers. When a cache miss occurs for cache C[i], a cache bypass prediction mechanism predicts whether the requested data should be cached in C[i]. A cache bypass prediction can be made when the processor or a higher level cache issues a cache request, when a C[i] cache miss occurs, or when the cache request is received at the C[i+1] cache. The cache bypass prediction mechanism includes appropriate bypass prediction states that can be maintained at the C[i] cache side, the C[i+1] cache side, or both. The bypass prediction state can be maintained for each C[i] cache line, for each C[i+1] cache line, or both. The bypass prediction state can also be maintained for each memory address, or for each set of memory addresses that are mapped to the same bypass prediction state based on some hashing function (for example).
Generally, the advantage of selectively enabling and disabling L1 cache bypass at cache line granularity using history-based predictive mechanisms increases L1 hit ratios leading to increased processor performance. There may be other advantages as well, for example decreasing the required L1-L2 data transfer bandwidth.
In summary, one aspect of the invention provides a method of enabling and disabling cache bypass in a computer system with a cache hierarchy, the method comprising the steps of: providing an identifying arrangement for identifying cache bypass status with respect to at least one cache line; providing a transferring arrangement which is adapted to: transfer a cache line identified as cache bypass enabled to one or more higher level caches of the cache hierarchy, whereby a next higher level cache in the cache hierarchy is bypassed; and transfer a cache line identified as cache bypass disabled to one or more higher level caches of the cache hierarchy, whereby a next higher level cache in the cache hierarchy is not bypassed; and selectively enabling or disabling cache bypass with respect to at least one cache line based on historical cache access information.
Another aspect of the invention provides a system comprising: a cache hierarchy; a plurality of cache lines associated with the cache hierarchy; an identifying arrangement for identifying cache bypass status with respect to at least one cache line; a transferring arrangement which is adapted to: transfer a cache line identified as cache bypass enabled to one or more higher level caches of the cache hierarchy, whereby a next higher level cache in the cache hierarchy is bypassed; and transfer a cache line identified as cache bypass disabled to one or more higher level caches of the cache hierarchy, whereby a next higher level cache in the cache hierarchy is not bypassed; and an arrangement for selectively enabling or disabling cache bypass with respect to at least one cache line based on historical cache access information.
Furthermore, an additional aspect of the invention provides a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for enabling and disabling cache bypass in a computer system with a cache hierarchy, the method comprising the steps of: providing an identifying arrangement for identifying cache bypass status with respect to at least one cache line; providing a transferring arrangement which is adapted to: transfer a cache line identified as cache bypass enabled to one or more higher level caches of the cache hierarchy, whereby a next higher level cache in the cache hierarchy is bypassed; and transfer a cache line identified as cache bypass disabled to one or more higher level caches of the cache hierarchy, whereby a next higher level cache in the cache hierarchy is not bypassed; and selectively enabling or disabling cache bypass with respect to at least one cache line based on historical cache access information.
For a better understanding of the present invention, together with other and further features and advantages thereof, reference is made to the following description, taken in conjunction with the accompanying drawings, and the scope of the invention will be pointed out in the appended claims.
A first preferred embodiment will be described with respect to three caches at three levels in a cache hierarchy, C[i−1], C[i], and C[i+1], where it is understood that if i=1 then cache C[0] actually is made up of the processor registers.
The case in which i=1 is illustrated in
In more detail, LB and RL entries are maintained in the directory 322 of cache C[i] 320 as follows. For each cache line L that bypasses C[i], a line bypass entry LB(L) is inserted into the directory 322 for line L. This entry contains the tag of line L, together with a counter initialized to 0. Note that if cache C[i] is a K-way set associative cache, then line L, had it been stored in C[i] (rather than bypassing C[i]), would always age out of C[i] after K misses (to lines other than line L, assuming LRU replacement) mapping to the set in which line L would have been stored. If a method other than LRU replacement is used (for example, tree-LRU replacement, or pseudo-random replacement), then line L could age out in less than K misses. In a first embodiment of the present invention, however, the counter for LB(L) will be incremented on every miss (to a line other than line L) to the set in which line L maps, and will be removed from the directory 322 when the counter reaches a value of K. As described below, if there is a miss for line L, the LB(L) entry may be removed before this point. Next, a replaced line entry RL(L′,L) is maintained in the directory 322 of cache C[i] 320 for each line L′ residing in the cache which replaced a line L when it was loaded in the cache. If line L′ is referenced before there is a miss to line L in cache C[i], then the RL(L′,L) entry will be removed from the directory. The RL(L′,L) entry will also be removed when line L′ ages out of the cache C[i], and (as described in detail below) if there is a miss to line L. Finally, bypass flags BF(L) are maintained in directory 335 as part of the cache directory entry 340 for each line L in cache C[i+1] 330. As described below, the bypass flag entries determine whether cache lines residing in C[i+1] 330 bypass cache C[i] 320 in the cache hierarchy. When a new line L is loaded into cache C[i+1] (due to a miss for L in C[i+1]), the bypass flag BF(L) for L may be initialized to 0 (bypass disabled for line L) or 1 (bypass enabled for L). In a first embodiment BF(L) will be initialized to 0 (bypass initially disabled for each new cache line L loaded into C[i+1]).
Given the preceding, the operation of a first embodiment of the present invention can now be described with reference to
Within the scope of the present invention, a number of design alternatives (or variant embodiments) are possible. For example, since the LB(L) and RL(L′,L) entries are used as predictors, an implementation alternative is store a subset of the tag bits for the cache line L in these entries. In this case a match of the corresponding subset of tag bits for line L (as in
As yet another example, rather than setting the bypass flag for a given cache line L using replaced line entries (as in the control flow shown in
Yet another variation would be to set the bypass flag BF(L) for line L only after M consecutive instances of no re-use of line L in cache C[i]. In more detail, a bypass counter BC(L) is stored in the directory entry for each line L residing in cache C[i+1], initialized to 0. Each time a given line L ages out of cache C[i] without being re-referenced in C[i], the counter BC(L) is incremented. If BC(L) is incremented and is then equal to M (where M is greater than or equal to 1), then BF(L) is set. Conversely, if a given line L residing in cache C[i] is re-referenced, then the counter BC(L) is set to 0. The resulting mechanism is a threshold method in which M consecutive instances of no re-use in cache C[i] are required to enable cache bypass. An important principle of these and other conceptually similar methods for setting cache bypass for a given line L is to predict, using usage history for the line L, whether cache performance would be improved if cache bypass were enabled for the given line L. In the case of the use of replaced line entries RL(L,L′), in some cases it is determined that loading L caused a miss to a line L′ that would not have occurred had line L bypassed the cache, and that there was no prior cache hit to line L; therefore the cache performance would have been improved if cache bypass had been enabled for line L. In the case of methods that determine whether line L is subsequently re-referenced, the improved performance (in the case that the prediction is accurate) rely on avoiding loading lines that are not subsequently re-used, effectively increasing the cache size and avoiding unnecessary traffic between levels of the cache hierarchy. In this context, threshold methods may be used to improve the accuracy of prediction, that is, a repeated pattern of no re-use is required to enable cache bypass for a given line.
A threshold method can also be used for determining when to disable cache bypass for a given line L which currently has cache bypass enabled (that is, BF(L) is set for the line L). In this method, a miss counter MC(L) is stored in the directory entry for each line L residing in cache C[i+1], initialized to 0 when the bypass flag for line L, BF(L), is set. On each miss for line L occurring in C[i] for which there exists a line bypass entry LB(L), the miss counter MC(L) for line L is incremented. If the counter MC(L) reaches a value of N (where N is greater than or equal to 1), then the bypass flag for L, BF(L), is cleared, and the counter MC(L) is re-set to 0. The principle of these and other conceptually similar methods is to clear cache bypass for a given line which would have been referenced in the next higher-level cache, or for the case in which N is 2 or greater, in which there is a repeated pattern of such misses, had cache bypass not been enabled for the given line.
Other conceivable variants for determining when to disable cache bypass for a given line L which currently has cache bypass enabled (that is, BF(L) is set for the line L) include methods which rely on recency information. For example, if there is a cache miss for line L in C[i] with BF(L) set in C[i+1] and line L is found to be in the MRU (most recently used) position in the set to which it maps in cache C[i+1] (assuming C[i+1] is configured as a set-associative cache with least recently used replacement within each set), then the bypass flag for line L, BF(L), is cleared. As is known, recency of usage can also be determined using counter-based methods. An important principle of these and other conceptually similar methods is to clear cache bypass for lines which are repeatedly requested from C[i+1] within short intervals of time, since such lines would more advantageously have been cached in the next higher level of the cache hierarchy.
It is to be understood that the present invention, in accordance with at least one presently preferred embodiment, includes: an identifying arrangement for identifying cache bypass status; a transferring arrangement which is adapted to: transfer a cache line identified as cache bypass enabled to one or more higher level caches of said cache hierarchy, whereby a next higher level cache in the cache hierarchy is bypassed; and transfer a cache line identified as cache bypass disabled to one or more higher level caches of said cache hierarchy, whereby a next higher level cache in the cache hierarchy is not bypassed; and an arrangement for selectively enabling or disabling cache bypass with respect to at least one cache line based on historical cache access information. Together, these elements may be implemented on at least one general-purpose computer running suitable software programs. These may also be implemented on at least one Integrated Circuit or part of at least one Integrated Circuit. Thus, it is to be understood that the invention may be implemented in hardware, software, or a combination of both.
If not otherwise stated herein, it is to be assumed that all patents, patent applications, patent publications and other publications (including web-based publications) mentioned and cited herein are hereby fully incorporated by reference herein as if set forth in their entirety herein.
Although illustrative embodiments of the present invention have been described herein with reference to the accompanying drawings, it is to be understood that the invention is not limited to those precise embodiments, and that various other changes and modifications may be affected therein by one skilled in the art without departing from the scope or spirit of the invention.
This invention was made with Government support under Contract No.: NBCH30390004 awarded by the Defense Advanced Research Projects Agency. The Government has certain rights in this invention.
Number | Name | Date | Kind |
---|---|---|---|
5625793 | Mirza | Apr 1997 | A |
5729713 | Leyrer | Mar 1998 | A |
6356980 | Arimilli et al. | Mar 2002 | B1 |
6647466 | Steely, Jr. | Nov 2003 | B2 |
Number | Date | Country | |
---|---|---|---|
20060112233 A1 | May 2006 | US |