Compiler Assisted Cache Prefetch Using Procedure Call Hierarchy

Microprocessor performance has been increasing at an exponential rate while memory system performance improved at a linear rate. This widening difference in performances is increasingly rendering advances in computer architecture less useful as more instructions spend more time waiting for data to be fetched from the memory after a cache miss. Data prefetching is a technique that avoids some cache misses by bringing data into the cache before it is actually needed. Different approaches to data prefetching have been developed, however existing prefetch schemes do not eliminate all cache misses and even with smaller cache miss ratio, miss latency remains an important performance limiter. In this thesis, we propose a technique called Compiler Assisted Cache Prefetch Using Procedure Call Hierarchy (CAPPH). It is a hardware-software prefetch technique that uses a compiler to provide information pertaining to data structure layout, data-flow and procedure-call hierarchy of the program to a mechanism that prefetches linked data structures (LDS). It can prefetch data for procedures even before they are called by using this statically generated information. It is also capable of issuing prefetches for recursive functions that access LDS and arbitrary access sequences which are otherwise difficult to prefetch. The scheme is simulated using RSIML, a SPARC v8 simulator. Benchmarks em3d, health and mst from the Olden suite were used. The scheme was compared with an otherwise identical system with no prefetch and one using sequential prefetch. Simulations were performed to measure CAPPH performance and the decrease in the miss ratio of loads accessing LDS. Statistics of individual loads were collected, and accuracy, coverage and timeliness were measured against varying cache size and latency. Results from individual loads accessing linked data structures show considerable decrease in their miss ratios and average access times. CAPPH is found to be more accurate than sequential prefetch. The coverage and timeliness are lower in CAPPH than in sequential prefetch. We suggest heuristics to further enhance the effectiveness of the prefetch technique