Technical Papers Archive

Faithful Prediction of the Performance of rocHPL Despite Hardware and Software Complexity


Authors: Jean Conan (Bull, French Institute for Research in Computer Science and Automation (INRIA)); Abdou Guermouche (French Institute for Research in Computer Science and Automation (INRIA)); Arnaud Legrand (CNRS, LIG); and Louis Poirel (Bull)

Abstract: High Performance LINPACK (HPL) remains the primary benchmark for ranking supercomputers and often serves as a dimensioning design component for system constructors. Its performance depends on a wide variety of parameters and remains difficult to predict on large, modern, heterogeneous clusters. Using stochastic performance models of computational kernels and the network simulation capabilities of SimGrid, a simulation framework for parallel application performance prediction, we design a lightweight simulator of the implementation of HPL on AMD GPUs (rocHPL). Taking advantage of the structure of HPL, we reduce the simulation complexity by several orders of magnitude. We evaluate the predictive capability of our approach across four distinct accelerated platforms comprising up to 1,024 GPUs, two software releases, and a wide range of parameters, and demonstrate its accuracy (simulated and real executions differ by 4% on average), flexibility, and scalability (at this scale, simulations take less than an hour on a laptop).


Presentation: file

Back to Technical Papers Archive Listing