Breiter, S. (2022):

Evaluating Sector Caches in High-Performance Computing

The sector cache is a hardware cache partitioning mechanism of the A64FX processor. The A64FX is used in the Fugaku system – currently the fastest supercomputer on the TOP500 list (as of November 2021). It allows application software to dynamically partition a cache and can reduce the occurrence of cache misses by protecting data with high temporal locality from eviction. Many cache partitioning techniques focus on optimizing the cache behavior of shared caches when multiple co-scheduled processes run on the same processor by assigning them to partitions. In contrast, the sector cache aims to improve the cache behavior of a single application by assigning its data to partitions. However, even the hardware man- ufacturer of the A64FX states that it is difficult to use the sector cache in a meaningful way. Therefore, a profiling tool based on the reuse distance metric is being developed using Intel’s PIN binary instrumentation framework. The profiling tool tries to provide program- mers with opportunities where the sector cache can be usefully applied without requiring the programmer to have detailed knowledge of a program’s data locality. Using the parallel NAS benchmarks as an example, it is shown that the tool can indeed help programmers to find code regions where the sector cache can improve cache behaviour. In addition, it is shown that sector cache can significantly improve performance in certain typical situations and these as well as the sector cache behavior of the A64FX are explored and analyzed.

• PDF Version (918.3 KB)