Speaker
Dr
George Bosilca
Description
Tracing a task-based application is necessary to get an idea of what's going on, but a heavy-handed utility could be so expensive that the trace might tell a story that does not look close to what the hardware is really doing when the trace is off. Using DPLASMA and PaRSEC, we demonstrate with GEMM some of the memory patterns on local ICL machines, which result from code written to visualize PaRSEC's native tracing abilities. We provide graphs along with takeaways regarding what this knowledge of memory might gain developers seeking to maximize performance on a task-based application.