A hybrid data prefetching architecture for data-access efficiency

High-performance computing has crossed the Petaflop mark and has been moving forward to reach the Exaflop range. However, while computing resources are making rapid progress, there is a significant gap between processing capacity and data-access performance. Due to this gap, although processing resources are available, they have to stay idle waiting for data to arrive, which has a severe impact on the overall system performance. In the meantime, applications tend to be more and more data intensive. The data-access delay, not the processor speed, has become the bottleneck of computing, especially for high-performance and high-end computing where performance is keen. There is a great need for research in improving data-access performance.;In this dissertation, we propose to improve data-access efficiency with a Hybrid Adaptive Prefetching architecture and associated innovative data prefetching techniques. The Hybrid Adaptive Prefetching architecture is built upon the memory hierarchy model, the for see engineering choice for masking the gap between computing and data-access speed, and enhances it with a hierarchical prefetching model to further mitigate the performance disparity and improve data-access speed. The fundamental idea behind the proposed solution is utilizing the excessive transistors on chip and available computing capability to build up specialized hardware and software approaches to accelerating data accesses, and thus to achieve a high sustained performance instead of a high peak performance. The Hybrid Adaptive Prefetching architecture reduces data-access latency via two stages, cache-memory stage by leveraging specialized hardware solutions and memory-disk stage by exploiting innovative software solutions. It improves data-access efficiency by harvesting the benefits of comprehensive, aggressive and adaptive prefetching strategies. The goal of this dissertation is to exploit hardware, compiler and system support to provide a systematic solution to boosting data-access performance for high-performance and high-end computing. Extensive experimental testing has been conducted to validate the design and verify the performance gain, and the results have demonstrated significant performance improvement. The Hybrid Adaptive Prefetching architecture can benefit a variety of applications such as scientific simulation, data mining, geographical information system, multimedia and visualization applications, etc. It will have a broad impact on improving data-access efficiency for high-performance and high-end computing.