Understanding the Scaling Behavior of the GFDL FMS High-Resolution Atmosphere Model on the Argonne BG/Q Platform IBM BG/Q: A Platform for Performance Discovery

We perform regression modeling to obtain estimates of the number of cycles (used as a proxy for application performance). A linear regression is performed in which the response (number of cycles) is a weighted sum of predictor variables (such as L1P misses, AXU instructions, XU instructions, FP operations, etc.) plus some random noise. Since linear relationships may not always be adequate to model the relationship between the predictors and response, non-linear modeling using cubic splines [Wah90] have also been performed. Both the linear and non-linear models are evaluated on blind test data using the R-squared and Mean Square Error (MSE) metric. Our results indicate that non-linear models are much better (higher R-squared) in capturing the characteristics of application performance modeling.