Multi-Level Execution Trace Based Lock Contention Analysis

Abstract

multi-threaded programming is a near-universal architecture in modern computer systems. Thread based programs usually utilize locks to coordinate access to shared resources. However, contention for locks can reduce parallel efficiency and degrade scalability.In this paper, we propose an execution-trace based method to analyze lock contention problems, without requiring an application’s source code. Our methodology uses dynamic analysis through execution tracing, running in several levels of the system to collect detailed runtime data. We combine it with an extended critical path algorithm which allows us to identify locking issues occurring in userspace. The result is a framework that is able to diagnose all contention issues while adding minimal impact on the system. We propose new views and structures to model and visualize collected data, giving programmers powerful comprehension tools to address contention issues.