Productive Data Locality Optimizations in Distributed Memory

Student: Engin Kayraklioglu (George Washington University)
Advisor: Tarek El-Ghazawi (George Washington University)

Abstract: With deepening memory hierarchies in HPC systems, the challenge of managing data locality gains more importance. Coincidentally, increasing ubiquity of HPC systems and wider range of disciplines utilizing HPC introduce more programmers to the HPC community. Given these two trends, it is imperative to have scalable and productive ways to manage data locality.

In this research, we address the problem in multiple ways. We propose a novel language feature that programmers can use to transform shared memory applications to distributed memory applications easily. We introduce a high-level profiling tool to help understand how distributed arrays are used in an application. As next steps, we are designing a model to describe the implementation of data locality optimizations as an engineering process, which can lend itself to combinatorial optimization. We are also implementing a profile-based automatic optimization framework that utilizes AI to replace the programmer completely in implementing optimizations for distributed memory.

Summary: pdf
Thesis Canvas: pdf



Back to Doctoral Showcase Archive Listing