Exploiting cache locality at run-time

Abstract

With the increasing gap between the speeds of the processor and memory system, memory access has become a major performance bottleneck in modern computer systems. Recently, Symmetric Multi-Processor (SMP) systems have emerged as a major class of high-performance platforms. Improving the memory performance of Parallel applications with dynamic memory-access patterns on Symmetric Multi-Processors (SMP) is a hard problem. The solution to this problem is critical to the successful use of the SMP systems because dynamic memory-access patterns occur in many real-world applications. This dissertation is aimed at solving this problem. Based on a rigorous analysis of cache-locality optimization, we propose a memory-layout oriented run-time technique to exploit the cache locality of parallel loops. Our technique have been implemented in a run-time system. Using simulation and measurement, we have shown our run-time approach can achieve comparable performance with compiler optimizations for those regular applications, whose load balance and cache locality can be well optimized by tiling and other program transformations. However, our approach was shown to improve significantly the memory performance for applications with dynamic memory-access patterns. Such applications are usually hard to optimize with static compiler optimizations. Several contributions are made in this dissertation. We present models to characterize the complexity and present a solution framework for optimizing cache locality. We present an effective estimation technique for memory-access patterns to support efficient locality optimizations and information integration. We present a memory-layout oriented run-time technique for locality optimization. We present efficient scheduling algorithms to trade off locality and load imbalance. We provide a detailed performance evaluation of the run-time technique.