An Application-Specific Microbenchmark for Memory Access

Benchmarking high performance computing systems is crucial to optimize memory consumption and maximize the performance of scientific application codes. We propose a configurable microbenchmark that explores the variations in memory bandwidth for a range of working set sizes, access patterns, and thread configurations. This framework is validated with the comparison of results from STREAM benchmark for repeated execution of the DAXPY triad kernel for both static and dynamic memory allocation. The access patterns emulate the common patterns found in simulation and modeling applications. Using application-specific access patterns we are able to refine the general roofline model for the target application