Study of the impact of non-determinism on numerical reproducibility and debugging at the exascale
Date
2017
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Non-determinism in high performance scientific applications has severe detri- mental impacts for both numerical reproducibility and accuracy, and debugging. As scientific simulations are migrated to extreme-scale platforms, the increase in platform concurrency and the attendant increase in non-determinism is likely to exacerbate both of these problems. In this thesis, we address the dual challenges of non-determinism’s impact on numerical reproducibility and on debugging. ☐ To address the numerical challenge, our work investigates the power of mathe- matical methods to mitigate error propagation at the exascale. We focus on floating- point error accumulation over global summations where enforcing any reduction order is expensive or impossible. We model parallel summations with reduction trees and identify those parameters that can be used to estimate the reduction’s sensitivity to variability in the reduction tree. We assess the impact of these parameters on the abil- ity of different reduction methods to successfully mitigate errors. Our results illustrate the pressing need for intelligent runtime selection of reduction operators that ensure a given degree of reproducible accuracy. ☐ To address the debugging challenge, our work examines the impact of logical clock ticking policies on the Clock-Delta Compression record-and-replay technique. We assess three logical clock ticking policies in terms of the number of out-of-order messages that result during recording executions under these policies. We examine the performance of Clock-Delta Compression when using the three ticking policies in four distinct application scenarios to probe the impact of floating-point workload and communication intensity on recording performance. Our results illustrate the pressing need for fine-grained logical clock ticking policies that reduce then out-of-order message rate of the Clock-Delta Compression record-and-replay technique.