Heartbeat Diagnosis of Performance Anomaly in OpenMP Multi-Threaded Systems

TL;DR

This paper introduces a heartbeat-based framework for diagnosing performance anomalies in OpenMP multi-threaded applications, utilizing injected APIs, feature extraction, and a tree-based algorithm to improve anomaly detection accuracy.

Contribution

It presents a novel heartbeat diagnosis framework with injected APIs and a tree-based algorithm for effective anomaly detection in OpenMP applications.

Findings

Successfully detects performance anomalies in benchmarks

Effective feature extraction from heartbeat sequences

Improves diagnosis accuracy in multi-threaded systems

Abstract

This paper presents a novel heartbeat diagnosis regarding performance anomaly for OpenMP multi-threaded applications. First, we design injected heartbeat APIs for OpenMP multi-threaded applications. Then, we leverage the heartbeat sequences to extract features of previously-observed anomalies. Finally, we adopt a tree-based algorithm, namely HSA, to identify the features that are required to diagnose anomalies. To evaluate our framework, the NAS Parallel NPB benchmark, EPCC OpenMP micro-benchmark suite, and Jacobi benchmark are used to test the performance of our approach proposed.