Profiling for detecting performance anomalies in concurrent software
Faheem Ullah, Thomas Groß · 2015
Understanding and identifying performance problems is difficult for parallel applications, but is an essential part of software development for parallel systems. In addition to the same problems that exist when analysing sequential programs, software development tools for parallel systems must handle the large number of execution engines (cores) that result in different (possibly non-deterministic) schedules for different executions. Understanding where exactly a concurrent program spends its time (esp. if some aspects of the program paths depend on input data) is the first step towards improving program quality. State-of-the-art profilers, however, aid developers in performance diagnosis by providing hotness information at the level of a class or method (function) and usually report data for just a single program execution. This paper presents a profiling and analysis technique that consolidates execution information for multiple program executions. Currently, our tool's focus is on execution time (CPU cycles) but other metrics (stall cycles for functional units, cache miss rates, etc) are possible, provided such data can be obtained from the processor's monitoring unit. To detect the location of performance anomalies that are worth addressing, the average amount of time spent inside a code block, along with the statistical range of the minimum and maximum amount of time spent, is taken into account. The technique identifies performance bottlenecks at the fine-grained level of a basic block. It can indicate the probability of such a performance bottleneck appearing during actual program executions. The technique utilises profiling information across a range of inputs and tries to induce performance bottlenecks by delaying random memory accesses. The approach is evaluated by performing experiments on the data compression tool pbzip2, the multi-threaded download accelerator axel, the open source security scanner Nmap and Apache httpd web server. An experimental evaluation shows the tool to be effective in detecting performance bottlenecks at the level of a basic block. Modifications in the block that is identified by the tool result in performance improvement of over 2.6x in one case, compared to the original version of the program. The performance overhead incurred by the tool is a reasonable 2-7x in majority of the cases.