Computer System With Multi-Buffer Data Cache for Prefetching Data Having Different Temporal and Spatial Localities

This application is related to U.S. patent application Ser. No. 07/526,341 filed May 18, 1990, now abandoned in favor of U.S. patent application Ser. No. 08/226,163, filed Apr. 8, 1994, that issued Dec. 5, 1995 as U.S. Pat. No. 5,473,764, and also to U.S. patent application Ser. No. 07/605,212 filed Oct. 24, 1990, now abandoned in favor of U.S. patent application Ser. No. 08/080,438 filed Jun. 21, 1993, which applications are incorporated herein by reference, as background material.

The invention relates to electronic computer systems. More particularly, the invention relates to a computer architecture which permits an efficient exchange of data between a relatively fast processor and a relatively slow memory.

Rapid advances in integrated circuit technology and in computer architecture have resulted in an increasing "memory reference delay gap" between relatively fast processing units and relatively slow memory. High performance processor chips with maximum throughput over 100 million instructions per second (MIPS) are available as commercial products. Multiprocessor systems using these processor chips as their processing units can also be implemented. However, the memory system needed to allow these high performance systems to run at their peak speed is either too complex or too expensive to be built. As a result, the performance bottleneck of many current high performance computer systems may not be the speed of their processing units, but rather is the efficiency of the memory system.

In order to overcome the problem of increasingly large memory reference delay gap between the fast processing units and slow memory, the prior art teaches using a the cache memory. Cache memory is a small high speed buffer memory between the central processing unit and the main memory of a computer system. Its primary purpose is to obtain high speed data/instruction accesses without the associated cost of building the entire primary memory using high speed technology. This is achieved by keeping data/instructions that are expected to be referenced in the near future in cache. In current cache designs, instruction caches are always given a higher priority to be implemented and optimized than data caches. If only a limited amount of chip area is available, it will be used to implement an instruction cache. Very often, a data cache is put onto the processor chip only after the instruction cache has been implemented and if there is still space available. This is due to the current level of understanding of instruction reference behavior and data reference behavior and to the accuracy of current cache models for instructions and for data. Since instruction references have a strong sequential reference characteristic, the prior art teaches a technique known as "cache prefetching" in which references are brought into the cache memory before they are actually needed. If the prefetching is correct, memory reference delay times can be overlapped (partially or completely) with program execution. For non-sequential instruction references due to branch or jump instructions, the prior art teaches a technique known as "branch target prediction" to guess which way the program execution is to go. Together with the cache prefetching technique, the number of cache misses can be reduced if the guessing is correct. Furthermore, a "branch target buffer or cache" can also be used to improve the cache hit ratio by placing branch targets into one buffer/cache and placing non-branch references into another buffer/cache. Consequently, the number of cache misses visible to the processing unit can be reduced and good instruction cache performance can be obtained. As used herein, the term "cache hit ratio" is defined as the probability that a data item requested by a processor unit will be found in cache memory and the "cache miss penalty" is defined as the time that the processing unit has to wait for the requested data item to arrive when a cache miss occurs.

In the prior art, data reference behavior is generally considered "random" and there is no accurate model for the prior art data cache designs. As a result, those caching techniques that can improve instruction cache performance might become ineffective when they are applied to data caches. For example, cache prefetching techniques are relatively much less efficient in data caches than in instruction caches. Sometimes, they actually decrease the machine performance. This is mainly due to problems of cache pollution and bus contention. Herein, "cache pollution" refers to the situation where data or instructions are fetched into the cache but are not referenced during the time they are in the cache. "Bus contention" is a situation where there are more requests trying to use the data bus than the bus can simultaneously serve. Compared to the behavior of instruction references, which are usually sequential in nature, the overall behavior of data references is more or less random. Since, in the prior art, the reference behavior of data has been much less predictable than the reference behavior of instructions, cache pollution occurs much more seriously in data cache prefetching than in instruction cache prefetching. Cache space is not only wasted to store prefetched non-referenced data, but data in the cache that are going to be referenced shortly might be replaced by the non-referenced data. Fetching of non-referenced data also creates bus traffic. In systems where bus bandwidth is a limited resource, bus contention might occur and machine performance will be further deteriorated. If the bus is busy prefetching non-referenced data, other requests to use the bus (e.g. instruction fetching) might be delayed. This occurs very often in high performance processor chips where the processing speed can be very fast but the bus bandwidth is limited by its physical constraints, for example, the total number of pins on a chip. Branch technique prediction and the use of branch target buffer/cache are not applicable to data caches. In the prior art data cache designs, data references are not divided into different types according to its sequential reference behavior. The distinction of branch references from sequential references in instruction references does not occur in data references. Thus, the performance of current data cache designs is much less than the performance of instruction cache designs.

Since the introduction of the cache in the 1960's, cache partitioning and multi-buffer cache techniques are often used to improve the cache hit ratio and to reduce the bus traffic. Some of the most commonly used prior art cache partitioning techniques are:

DATA CACHE VS. INSTRUCTION CACHE--the cache space is divided into two mutually exclusive partitions, one for instruction storage and one for data storage.

CACHE PARTITION FOR REAL-TIME SYSTEMS--the cache space is divided into several partitions, only one of which will be assigned and used by a task each time the task is running.

ADDITION OF BRANCH TARGET CACHE--a separate buffer/cache, used to store only branch target instructions, is added onto the existing cache.

ADDITION OF PREFETCH BUFFER--a separate buffer, used to store only sequentially prefetched instructions, is added onto the existing cache.

ADDITION OF VICTIM CACHE or (or"MISS CACHE") a separate buffer/cache, used to temporarily save data that are just replaced from the cache, is added between the existing cache and the main memory.

All these techniques lose their effectiveness in data cache designs. The first technique (data cache vs. instruction cache) tries to avoid the interference of instruction references from data references in caches while the second technique (cache partitioning for real-time systems) tries to avoid the interference of instructions and data references of one task from instructions and data references from the other tasks. However, the problem of unpredictable data reference behavior remains. The third techniques (addition of branch target cache) cannot be applied in data caches because there are no branch targets in data references. The fourth technique (addition of prefetch buffer) is not very useful in improving data cache performance because the reference behavior of data is more or less random rather than sequential. Sequential prefetching in data references often deteriorates data cache performance instead of improving it. The last technique (addition of victim cache or miss cache) can improve the data cache performance slightly by adding more associativity to the cache. However, it still does not solve the problem of unpredictable data reference behavior.

A further discussion of the organization and construction of cache memory buffers and systems is found in "High-Speed Memory Systems", Pohm and Agrawal, Reston Publishing Company (1983) and in "Cache Memories", by A. J. Smith, Computing Surveys, Volume 14, Number 3, September 1982, pp. 473-530, which are incorporated herein, by reference, as background material.

Data references are divided into two types called S.sub.-- References and P.sub.-- References according to the addressing mode and the registers used in the address calculation. Data references from main memory are stored in a cache which comprises two buffers, an S.sub.-- Buffer and a P.sub.-- Buffer, each of which is optimized and used to handle one type of data reference. The S.sub.-- Buffer is optimized for S.sub.-- References and the P.sub.-- Buffer is optimized for P.sub.-- References. Each of these two types of references has its own distinct characteristics that are not shared by the other reference type. Thus, a significant improvement in cache performance is obtained with the multi-buffer data cache.

In current cache designs, program reference behavior is often explained by the principle of locality of reference. There are two types of locality references, "temporal locality" (or locality of time) and "spatial locality" (or locality of space). Over a short period of time, memory references made by a reference program are not uniformly distributed over the address space. Instead, references often occur in small clusters--the working set of references. The first type of locality, temporal locality, refers to the fact that information currently being used is likely to be used again in the near future. This is the basic rationale for most prior art caches: most programs attempt to use data or code more than once during program execution. This type of reference behavior can be found in program loops in which both data and instruction references are expected to be used repeatedly for each loop or in subroutines that are called several times during the program execution.

The address space usually is grouped into a number of individual contiguous segments of that space and only a few segments will be referenced repeatedly at the same time. The second type of locality, spatial locality, suggests that the loci of references of a program in the near future will be likely to be near the current loci of reference. This type of reference behavior is common in scientific programs where instructions are mostly executed in a sequential order and it is common to access sequences of array elements, which are typically contiguous in memory.

With these localities of time and space, the present invention attempts to keep information which the processor is currently working on (local in time), together with information near the current loci of references (local in space), in cache and readily available. However, it is generally agreed that this design principle works well in instruction caches, but not in data caches. This is due to the very different reference behavior between instructions and data. While instruction reference behavior has a strong sequential reference property and is highly predictable, the prior art teaches that data reference behavior is random and is much less predictable than instruction references. Since most of the current cache models in management schemes are primarily designed for instruction caches, it is not surprising that these schemes and control mechanisms are much less effective when they are applied to data caches. However, careful study shows that data reference behavior is actually not as random as it looks. In fact, it is a combination of S.sub.-- References and P.sub.-- References, and each of these two types of references is highly predictable. The first type of data references, the P.sub.-- Reference, is mainly due to accesses of data array elements using pointers. It has strong sequential reference properties that are similar to the sequential reference property of instructions. Hence, the next reference of this type of data is highly predictable. That is, it possesses a strong forward (or perhaps backward) spatial locality of references. However, life span of cached P.sub.-- References is usually very short, usually a few instructions, and the reference frequency of P.sub.-- references is very low. This means that the temporal locality of P.sub.-- References is very weak.

The second type of data references, the S.sub.-- References, is mainly due to accesses of stack variables and global scalar variables. This type of references is usually referenced very often during their live range period. Thus, a strong temporal locality of reference property can be found in this type of references. However, the spatial locality of S.sub.-- References is very weak because its access pattern is almost random.

By inspecting the addressing mode and the registers used in the address calculation during instruction decoding, data references can be ascertained to be S.sub.-- References or P.sub.-- References at run time with high accuracy. This is because only references using indirect addressing mode might be P.sub.-- References and some registers are often designed for some specific purposes. For example, references using a direct addressing mode are likely to be S.sub.-- References; indirect references using the stack pointer register are probably local stack variables.

By dividing data references into S.sub.-- References and P.sub.-- References, the locality of each of these two types of data reference can be described more accurately. Based on this data reference model, a computer system of the present invention includes a data cache having an S.sub.-- Buffer which is optimized for prefetching and storing S.sub.-- Reference data and a P.sub.-- Buffer which is optimized for prefetching and storing P.sub.-- Reference data. The data cache includes a control unit which analyzes the addressing mode and registers used in the address calculation during instruction decoding to determine if the reference data should be an S.sub.-- Reference or a P.sub.-- Reference.

The invention will be described by reference to FIG. 1, the lone drawing FIGURE which illustrates a computer system with a data cache of the present invention.

The FIGURE is a multi-buffer data cache of the present invention. The cache controls the fetching of data which is referenced by a processing unit 10 from a main memory 20. As is usual in the prior art, the processing unit includes a decode unit 30 which decodes instructions to determine, among other things, the types of data references contained therein and the registers used for indirect mode addressing.

A multi-buffer data cache 100 is connected between the main memory 20 and the processing unit 10. It consists of an S.sub.-- Buffer 40, a P.sub.-- Buffer 50, and a control unit 60. The control unit 60 regulates the flow of data so that each of the P.sub.-- Buffer and S.sub.-- Buffer has its own distinct unique features and control schemes to optimize one type of data reference: the S.sub.-- Buffer is used to handle S.sub.-- References and the P.sub.-- Buffer is used to handle P.sub.-- References. The buffers are otherwise constructed and connected in the manner of the prior art.

Whenever a data reference is found in some instruction, its addressing mode and the registers used in the address calculation are determined by the decode unit 30 and sent to the control unit 60 which determines whether it is an S.sub.-- Reference or a P.sub.-- Reference. This is done according to predefined rules which are incorporated in the control unit logic. For example, a reference that uses indirect addressing mode and uses the stack pointer register in its address calculation is considered to be an S.sub.-- Reference. When a data reference is made, each of the two buffers is checked to determine whether there is a cache miss or a cache hit.

The S.sub.-- Buffer is mainly used to store S.sub.-- References. Since an S.sub.-- Reference does not have a strong sequential reference priority, no sequential prefetching is used. This is because the serious cache pollution that data prefetching generates might result in performance deterioration instead of performance improvement. This is especially true if the S.sub.-- Buffer size is small. When there is an S.sub.-- Buffer cache hit on an S.sub.-- Reference, there is no change in the content of the cache. When there is an S.sub.-- Buffer cache hit for a P.sub.-- Reference, there is no change in the content of the cache. When there is a cache miss for an S.sub.-- Reference, the cache block corresponding to that miss is fetched from main memory into the S.sub.-- Buffer.

The P.sub.-- Buffer is used mainly to store P.sub.-- References. Since there is a strong sequential reference property in P.sub.-- References, sequential prefetching is used in the P.sub.-- Buffer. When there is a P.sub.-- Buffer cache hit on a P.sub.-- Reference, there is no change in the content of the cache. When there is a P.sub.-- Buffer hit for an S.sub.-- Reference, the cache block corresponding to that reference is transferred from the P.sub.-- Buffer to the S.sub.-- Buffer. After the block transfer, the cache block in the P.sub.-- Buffer that contains the current S.sub.-- Reference is marked as empty. When there is a cache miss for a P.sub.-- Reference, the cache block corresponding to the miss is fetched from main memory into the P.sub.-- Buffer. All prefetched data is placed in the P.sub.-- Buffer. A prefetching scheme which allows constantly prefetching can be used.

Although the invention has been described with respect to a two-buffer cache, it is not so limited. Additional cache buffers may be provided, in addition to the P.sub.-- Buffer and the S.sub.-- Buffer, with specific data fetching and storage characteristics which are optimized for other identifiable data reference types which have unique characteristics.