Iceberg: a tool for static analysis of Java critical sections

Abstract

In this paper we present a static analysis tool called Iceberg that helps programmers find potential performance bugs in concurrent Java programs. The focus of our work is on identifying critical sections with high variability in their latency: in most cases they execute quickly, but occasionally they stall, holding a lock for an unusually long time, and preventing other threads from making progress. The end user experiences such behavior as transient program "hangs". These performance bugs are difficult to find because they are infrequent, transient, and hard to reproduce. This paper describes our initial results running Iceberg on 24 real-world concurrent programs. Our current approach is to identify all of the code that could be executed inside each critical section, including all methods potentially called by it. We collect a number of code metrics that might indicate potential performance problems. These metrics include counts of variable-latency operations such as I/O and memory allocation, as well as overall measurements of critical section size. Using our tool we are able to find critical sections with unusual behavior compared to the other critical sections. Our future work includes a more detailed analysis of control-flow through critical sections, as well as a dynamic analysis to measure the critical section latencies directly.