Comparing the performance of rigid, moldable and grid-shaped applications on failure-prone HPC platforms

Abstract This paper compares the performance of different approaches to tolerate failures for applications executing on large-scale failure-prone platforms. We study (i) R i g i d applications, which use a constant number of processors throughout execution; (ii) M o l d a b l e applications, which can use a different number of processors after each restart following a fail-stop error; and (iii) G r i d S h a p e d applications, which are moldable applications restricted to use rectangular processor grids (such as many dense linear algebra kernels). We start with checkpoint/restart, the de-facto standard approach. For each application type, we compute the optimal number of failures (i.e. that maximizes the yield of the application) to tolerate before relinquishing the current allocation and waiting until a new resource can be allocated, and we determine the optimal yield that can be achieved. For G r i d S h a p e d applications, we also investigate Application Based Fault Tolerance (ABFT) techniques and perform the same analysis, computing the optimal number of failures to tolerate and the associated yield. We instantiate our performance model with realistic applicative scenarios and make it publicly available for further usage. We show that using spare nodes grants a much better yield than currently used strategies that restart after each failure. Moreover, the yield is similar for R i g i d , M o l d a b l e and G r i d S h a p e d applications, while the optimal number of failures to tolerate is very high, even for a short wait time in between allocations. Finally, M o l d a b l e applications have the advantage to restart less frequently than R i g i d applications.