An End-to-End Example of Performance Modeling in Action: Hybrid Architectures and Applications
Abstract: We will begin with a brief description of performance activities at Los Alamos. Modeling work related to Roadrunner, the fastest machine currently in use and the first system that achieved a Petaflop/s sustained in Linpack, will be the emphasis of the presentation. In relation to Roadrunner, we will describe uses of performance modeling techniques to system and application design for heterogeneous/hybrid clusters, and to performance prediction for full applications on the Roadrunner architecture before the system was built. A case study of tool development guided by modeling will be presented, as it applies to programming models for accelerated clusters. We will conclude with some thoughts related to the performance and programmability of heterogeneous clusters for a broad spectrum of workload characteristics.