Performance Prediction for Scalability Analysis

Abstract

Performance prediction is an effective approach for understanding the scalability of large-scale parallel applications when the whole target systems are not available. However, accurate and efficient performance prediction is difficult because the execution time of parallel applications is determined by several factors, including sequential computation time in each process, communication time, and their convolution. This chapter proposes a novel approach to acquire the sequential computation time accurately and efficiently, which only needs a single node of the target platform. First, we employ deterministic replay techniques to execute any process of a parallel application on a single node at real speed. Thus, we can simply measure the real sequential computation time on a target node for each process one by one. Second, we observe that processes in parallel applications can be clustered into a few groups where processes in each group have similar computation behaviors. Based on this observation, we only execute representative parallel processes and significantly reduce measurement time. We implement a performance prediction framework, called Phantom, which integrates the above computation time acquisition approach with a trace-driven network simulator. We validate our approach on several platforms, and the prediction error of our approach is less than 8% on average. (Ⓒ 2015 IEEE. Reproduced, with permission, from Jidong Zhai, et al., Performance prediction for large-scale parallel applications using representative replay, IEEE Transactions on Computers, 2015.)