P93: Spacehog: Evaluating the Costs of Dedicating Resources to In Situ Analysis
Abstract: Using in situ analytics requires that computational resources be shared between the simulation and the analysis. With space-sharing, there is a possibility for contention over these shared resources such as memory, memory bandwidth, network bandwidth, or filesystem bandwidth. In our analysis, we explore the sensitivity of different applications with a set of microbenchmarks that are representative of analytics that may be used with scientific simulation. These tasks are modeled using a library called libspacehog. The experimentation consisted of examining three different dimensions of how simulation workloads might be space-shared with analysis codes. The results indicate that contention does need to be considered when applying in situ analytic techniques and can be of greater concern than simply the number of analysis processes or overall process density. This research provides an explanation on how the application’s performance is affected by space-sharing to further understand in situ analytic techniques.
Award: Best Poster Finalist (BP): no
Two-page extended abstract: pdf