Automatic Performance Prediction For Load-Balancing Coupled Models

Computationally-demanding, parallel coupled models are crucial to understanding many important multiphysics/multiscale phenomena. Load-balancing such simulations on large clusters is often done through off-line, static means that often require significant manual input. Dynamic, runtime load-balancing has been shown in our previous work to be effective, but we still used a manually generated performance predictor to guide the load-balancing decisions. In this paper, we show how timing and interaction information obtained by instrumenting the middleware can be used to automatically generate a performance predictor that relates the overall execution time to the execution time of each individual submodel. The performance predictor is evaluated through the new coupled model benchmark employing five constituent submodels that simulates the CCSM coupled climate model.