Authors: Jim Galarowicz and Martin Schulz

Open|SpeedShop, is an open source, multi platform Linux performance analysis tool for both sequential and parallel (threading and MPI) applications. It directly targets application developers and computer scientists as its main users. It provides easy access to an application’s performance profile, while still allowing more sophisticated and detailed analysis found in other complex, but more difficult to use tools. For the novice user, Open|SpeedShop provides convenience command scripts to collect the application performance data and provides default methods of viewing collected data. Expert users appreciate low-level access to performance data and scripting interfaces, which include Python integration. The tool set currently targets Linux clusters built using Intel, AMD, and IBM PowerPC processors and works on a wide variety of Linux distributions. Additionally, O|SS is available on specialized HPC platforms like IBM’s Blue Gene and Cray’s XT/XE/XK lines.

The performance analysis functionality in Open|SpeedShop includes a set of specific experiments that allow the user to easily gather a variety of different performance statistics about an application. This includes Program Counter (PC) sampling, a lightweight way to get an overview of application performance bottlenecks; Call Stack Sampling analysis, a technique to find hot call paths; Hardware Performance Counters, providing access to low level information such as cache or translation look-aside buffer (TLB) misses; Message Passing Interface (MPI) Profiling and Tracing, enabling users to detect MPI communication bottlenecks; I/O Profiling and Tracing to study an application’s I/O characteristics; and Floating Point Exception (FPE) analysis to detect floating point exceptions that can slow down applications.

Once the performance data is collected, it is stored in the form of a single relational database. Open|SpeedShop then reads the performance information from the database and displays the data through a set of detailed reports that allow the user to easily relate the performance information back to their application source code. This information is accessible through a graphical user interface (GUI), from a command line interface, as well as from within Python scripts. Additionally, the tool set includes a series of analysis techniques, including outlier detection, load balance analysis, and cross experiment comparisons. In summary, Open|SpeedShop functionality provides a comprehensive set of techniques that greatly aid in the analysis and understanding of parallel application performance.