A New Memory Benchmark for Modeling Irregular Applications

For some time now, data movement has been a first-order concern for developing high-performance code. Many benchmarks have been developed to measure memory performance, but none so far have accurately modeled the irregular data access patterns seen in programs like sparse linear solvers and graph algorithms. This makes the task of designing new memory systems to accelerate such irregular applications difficult. To remedy this, we design and implement a new benchmark based on scatter/gather access primitives, SGbench. In this talk, I will discuss the implementation of SGbench and its ability to help us measure the potential impact of new memory technologies like the Hybrid Memory Cube. I will discuss how this benchmark helps us model the memory access patterns in the sparse solver SuperLU_Dist.