Analysis of Application Sensitivity to System Performance Variability in a Dynamic Task Based Runtime

Application scalability can be significantly impacted by node level performance variability in HPC. While previous studies have demonstrated the impact of one source of variability, OS noise, in message passing runtimes, none have explored the impact on dynamically scheduled runtimes. In this pa-per we examine the impact that OS noise has on the Le-gion runtime. Our work shows that 2.5 % net performance variability at the node level can result in 25 % application slowdown for MPI+OpenACC based runtimes compared to 2 % slowdown for Legion. We then identify the mechanisms that contribute to better noise absorption in Legion, quan-tifying their impact. Furthermore, we assess the impact of OS noise at the granularity of communication, task schedul-ing, and application level tasks within Legion demonstrating that where noise is injected can significantly effect scalabil-ity. The implications of this study on OS and runtime ar-chitecture is then discussed. 1