Vadar: Runtime Performance Variance Detection and Diagnosis for Parallel Applications
TL;DR
Vadar adopts a non-intrusive dynamic interception mechanism to monitor parallel applications with state transition detection to cope with workload change at runtime, and provides comprehensive monitoring for both regular and irregular computational workloads.
Abstract
Performance variance is a common and serious issue for parallel applications in numerous scientific computing, data-mining and deep learning on modern computer systems, which can cause unexpected and unreproducible performance degradation. Therefore, the detection and diagnosis of performance variance is crucial for both computer systems and parallel applications. However, there are various workloads and workloads often change during the execution of parallel applications, so it is extremely difficult to detect and diagnose performance variance. In this paper, we propose Vadar, a runtime performance variance detection and diagnosis tool for parallel applications. Vadar adopts a non-intrusive dynamic interception mechanism to monitor parallel applications with state transition detection to cope with workload change at runtime. To reconstruct the execution logic, Vadar builds an asynchronous state transition graph comprising four types of dependency edges, which accurately captures sequential execution, inter-process communication, and CPU-GPU heterogeneous dependencies. Furthermore, by integrating a performance model to detect the performance variance of sparse matrix vector multiplication, Vadar provides comprehensive monitoring for both regular and irregular computational workloads. The evaluation results demonstrate that Vadar effectively detects performance variance in real applications with acceptable overhead and identifies the root-cause workload of performance variance.
