CSAP: A Performance Predictor for Climate Simulation Applications on Intel CPUs

摘要

The development of climate simulation applications highly depends on high-performance computers. A fundamental problem in constructing a high-performance computer is how to select the processors. A straightforward solution is to use a state-of-the-art processor version available from the market. However, this approach is not economical for climate simulation applications. To pick out a suitable processor version, we propose to predict the performances of climate simulation applications on various processor versions, and design and develop a new cross-architecture performance predictor CSAP (Climate Simulation Applications performance Predictor), based on the performance characteristics of climate simulation applications. Our experimental evaluation shows that CSAP can predict the performances of climate simulation applications on various Intel CPU versions with high accuracy and low overhead.