Performance, Power Modeling and Optimization for High-Performance Computing Systems

Abstract

Heterogeneity abounds in modern high-performance computing systems. Applications are heterogeneous, containing time-varying unbalanced utilization for different resources, and system architectures have become heterogeneous in order to achieve higher levels of performance and energy efficiency. The most powerful, and also the most energy-efficient high-performance computing systems today consist of many-core CPUs and GPGPUs with a variety of specialize on-chip and off-chip memories. These heterogeneous systems provide a huge amount of computing resources, but it is becoming increasingly challenging to use them effectively and efficiently to maximize their potential. This becomes an even more pressing challenge as energy efficiency becomes the primary barrier to achieving higher levels of performance. This thesis addresses the challenges of performance modeling and optimization in heterogeneous high-performance computing systems. Effective system optimization requires understanding of how performance and power change in response to optimizations. Therefore, we begin by summarizing the impact of modern architectural advances on performance and power modeling for chip multiprocessors (CMPs). We present two models that estimate the performance and power in such systems. The first model, CAMP, is a fast and accurate cache-aware performance model that estimates the performance degradation due to cache contention of processes running on cache-sharing cores. We then propose a system-level power model for a multi-programmed CMP environment that accounts for cache contention. We explain how to integrate the two models to enable power-aware process assignment. Then, we propose an off-chip memory access-aware runtime DVFS control technique that minimizes energy consumption subject to a constraint on application execution time. The second part of the dissertation focuses on improving performance for GPGPUs. After a thorough analysis on CPI breakdown, we lay out all the key factors that govern GPU throughput. In order to improve overall performance for GPGPUs, we propose two approaches that address the key factors, without introducing extra congestion and degradation to the system. We first propose a new two-level priority scheduling policy to improve overall performance by optimizing effective degree of parallelism. Then, we propose ICMT, a full, detailed solution for intra-core multitasking for GPGPUs, including architectural support and a contention-aware workload scheduling algorithm that improves all the key factors in a balanced fashion. Furthermore, we propose a new contention-aware analytical performance model that provides fine-grained workload scheduling decisions for intra-core multitasking, including detailed resource allocation from co-scheduled workloads.