Performance Dynamics of Massively Parallel Codes (LMAC)

Felix Wolf, German Research School for Simulation Sciences

(slides)

The deployment of adaptive algorithms and sophisticated load-balancing schemes make the execution behavior of simulation programs increasingly dynamic. Effective application optimization therefore requires capturing and analyzing performance data along the dimensions of both space and time. While existing performance analysis tools typically provide detailed information along spatial dimensions like processes and nodes, the aspect of performance dynamics has so far been neglected.

To support the optimization of applications with time-dependent performance characteristics, the BMBF project LMAC (Leistungsdynamik massiv-paralleler Codes) aims to extend the established performance-analysis tools Vampir, Scalasca, and Periscope with new functionality to measure and analyze performance dynamics. In addition, the University of Oregon, an associated partner in the project, complements these objectives with corresponding extensions to the performance tool TAU. In a broader interpretation of "performance dynamics", the project also targets performance differences across multiple program versions created in the development process. A major portion of the new capabilities will be integrated in Score-P, a measurement infrastructure shared by all of the above-mentioned tools, which was created in the BMBF project SILC (2009-2011).

To serve a broad user base both within the Gauss-Alliance and beyond, most of the performance tools are released free of charge to the community under an open-source license. Only Vampir, due to its sophisticated user interface, is distributed commercially. The software products are accompanied by training- and support offerings through the Virtual Institute – High Productivity Supercomputing, and will be maintained beyond the original project duration.