Performance prediction of massively parallel computation by Bayesian inference

Abstract

A performance prediction method for massively parallel computation is proposed. The method is based on performance modeling and Bayesian inference to predict elapsed time T as a function of the number of used nodes P (T = T (P )). The focus is on extrapolation for larger values of P from the perspective of application researchers. The proposed method has several improvements over the method developed in a previous paper, and application to realsymmetric generalized eigenvalue problem shows promising prediction results. The method is generalizable and applicable to many other computations.

Data-analysis software framework 2DMAT and its application to experimental measurements for two-dimensional material structures

“…In addition, users can implement and analyze any objective function F(X). For example, in a previous study [41], the REMC algorithm in 2DMAT was used as the performance prediction method for a massively parallel numerical library for a generalized eigenvalue problem. This paper shows a method by which to predict the elapsed time T for the numerical library as a function of the number of nodes used P (T = T (P)) and focuses on extrapolation to larger values of P. The teacher dataset is the set of measured elapsed times for various numbers of nodes {T exp (P i )} i=1,...,ν .…”

Automatic performance tuning using the ATMathCoreLib tool: Two experimental studies related to dense symmetric eigensolvers

“…Automatic performance tuning of matrix libraries has been studied from various aspects. There are approaches based on exhaustive search, 15,16 incremental parameter sampling, 17 statistical models 18,19 and machine learning, 20,21 to mention a few. Among them, the approach of ATMathCoreLib is unique in that it is targeted at the finite horizon problem; it is designed to finish auto-tuning in a specified number of executions and minimize the total execution time.…”