Performance tuning of distributed applications with CoSMoS

Performance monitoring of distributed software executing on workstation cluster systems is a complex task because of the multitude of factors that can affect performance in these environments. For performance analysis to be done effectively, all runtime events which have an impact on performance have to be captured. We introduce an integrated monitoring system that is capable of recording performance data on the application level, the operating system level, and the network level. It is thus able to provide complete coverage of all subsystems relevant to application performance. One of the key features of the proposed monitoring system is its ability to correlate performance data with application source code, pointing the developer to the portions of his code that may be time-critical. The system thus supports a cyclic process of optimizing performance in multiple iterations of analysis and refinement. This method allows performance bottlenecks to be tracked down to source code instruction level.