ASLOP: A field-access affinity-based structure data layout optimizer

Abstract

By rearranging the data, data layout optimizations improve the utilization of a cache line between two of its successive refills, thus reducing the total number of cache line refills and improving the performance of a program. In this paper, we show that to enable structure data layout optimizations to be effective, two parameters, namely intra-instance affinity and inter-instance affinity, need to be considered at the same time in order to model the cache line utilization more accurately. We also propose a lightweight approach to measure intra-instance affinity and inter-instance affinity to avoid complex memory trace analyses. A prototype, called ASLOP, has been implemented in the Open64 compiler and evaluated using benchmarks from SPEC CPU 2000, SPEC CPU 2006 and Olden benchmark suites that have extensive structure types. Our approach can achieve up to 48.1% performance improvement over the original programs, and 11.9% over the optimized programs using maximal reshaping, an existing approach that is known to produce close to the best results, on the two platforms we tested.

STrans: a comprehensive framework for structure transformation

“…As STrans is a part of ASLOP, all benchmarks evaluated in ASLOP [3] are well transformed by STrans, including 179.art and 181.mcf from SPEC CPU 2000 [9] , 462.libquantum, 429.mcf, and 472.moldyn from SPEC CPU 2006 [10] , em3d, health, treeadd, and tsp from Olden benchmarks [11] . For each benchmark, we have applied each transformation method.…”

“…There are many papers focusing on making decision on a good SDLO or general data layout strategy [2,3,[12][13][14][15][16] .…”

“…STrans is used as the transformer in ASLOP [3] , which makes SDLO decision based on both intra-instance affinity and inter-instance affinity.…”

“…Structure Data Layout Optimization (SDLO) is a prevailing technique to improve cache utility for structure fields access by compiler optimization. SDLO has been proved to be very effective for many applications, and we can usually get 20% to over 100% performance improvement for those applications [1][2][3] .…”

“…In this paper, we present STrans, the transformer in ASLOP [3] . We have made two major contributions:…”