In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy

Cheng Wang

will defend his dissertation proposal

Exploiting Data Locality for Irregular Applications on Shared­ Memory Multicore Architectures

Abstract

The last decade has witnessed a rapid increase of the gap between CPU speed and memory bandwidth as new generations of computer systems are introduced. With the rapid improvement of processor speed, the performance of the memory hierarchy has become the principal bottleneck for most applications. For data-parallel applications where locality is abundant, it is a relatively straightforward task to port and optimize for a specific architecture. A number of compiler transformation techniques have been developed to improve data locality for those regular applications, thus reducing the total number of main memory accesses in a program.

Yet to date, most of the data locality transformation techniques are static and can be only applied at compile time. As a result, these techniques cannot be used to optimize irregular and dynamic applications, in which the memory layout and data access pattern remain unknown until runtime and may even change dynamically during the computation. It is becoming increasingly important to exploit data locality for these irregular applications as sparse data structures with irregular memory access patterns have being identified in many important scientific and engineering applications. The work in this dissertation aims to make a fundamental contribution to improve data locality and exploit parallelism for irregular applications, both of which are essential for improving the applications' performance on state-of-the-art parallel computer systems. In particular, we propose an online transformation framework which reorders the computation and data layout at runtime. The framework effectively circumvents the inherent complexity in finding an optimal data layout, making it feasible to improve the data locality for a variety of irregular applications with minimum runtime overhead.

Date: Tuesday, May 5, 2015
Time: 10:30 AM
Place: PGH 218
Advisor: Prof. Barbara Chapman

Faculty, students, and the general public are invited.