Abstract

<p>In this thesis we propose and evaluate a fully automatic technique to manage the IO requirements of out of core scientific applications. In our scheme the compiler determines information on future access patterns and inserts prefetch and release operations to initiate IO before the data is needed and to free pages that are no longer accessed. No modifications to the original application source code are needed to support our compiler analysis. The operating system provides support for the basic prefetch and release operations and interacts with a run time layer to accelerate performance by minimizing prefetch overhead.</p> <p>&nbsp;</p> <p>We have implemented our scheme using the SUIF compiler and the Hurricane operating system. Our experimental results demonstrate that our fully automatic scheme effectively hides the IO latency in out of core versions of the entire NAS Parallel benchmark suite thus resulting in speedups of roughly twofold for five of the eight applications with one application speeding up by over threefold. </p>

 

Manuscript

 

Bibtex