P.P. Veeraraghavan and J.J. Evans (USA)
Parallel application run time sensitivity, communication performance evaluation
Over the past several years new commodity HPC systems have been deployed using state-of-the-art compute node ar chitectures. We have been studying parallel application run time sensitivity to network performance degradation, de veloping tools for characterizing application behavior and models and metrics for quantifying run time sensitivity on commodity single-core HPC systems. This work re examines the same problems, using the same tools and methods to determine whether multiple multi-core proces sor HPC systems can be characterized the same way in terms communication performance and application sensi tivity. The contributions of this work are assessments at the core, processor, and node level for communication param eter determination in terms of startup and transport time variables for common communication types. Another con tribution is a preliminary run time sensitivity assessment of several common HPC benchmarks.
Important Links:
Go Back
