Title:SPARK00: A Benchmark Package for the Compiler Evaluation of Irregular/Sparse Codes

Abstract: We propose a set of benchmarks that specifically targets a major cause of performance degradation in high performance computing platforms: irregular access patterns. These benchmarks are meant to be used to asses the performance of optimizing compilers on codes with a varying degree of irregular access. The irregularity caused by the use of pointers and indirection arrays are a major challenge for optimizing compilers. Codes containing such patterns are notoriously hard to optimize but they have a huge impact on the performance of modern architectures, which are under-utilized when encountering irregular memory accesses. In this paper, a set of benchmarks is described that explicitly measures the performance of kernels containing a variety of different access patterns found in real world applications. By offering a varying degree of complexity, we provide a platform for measuring the effectiveness of transformations. The difference in complexity stems from a difference in traversal patterns, the use of multiple indirections and control flow statements. The kernels used cover a variety of different access patterns, namely pointer traversals, indirection arrays, dynamic loop bounds and run-time dependent if-conditions. The kernels are small enough to be fully understood which makes this benchmark set very suitable for the evaluation of restructuring transformations.