Date of Award

12-2016

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering

First Advisor

Vijay Pai

Committee Chair

Vijay Pai

Committee Member 1

Milind Kulkarni

Committee Member 2

Samuel P. Midkiff

Committee Member 3

Arun Prakash

Abstract

With modern advancements in hardware and software technology scaling towards new limits, our compute machines are reaching new potentials to tackle more challenging problems. While the size and complexity of both the problems and solutions increases, the programming methodologies must remain at a level that can be understood by programmers and scientists alike. In our work, this problem is encountered when developing an optimized framework to best exploit the semantic properties of a finite-element solver. In efforts to address this problem, we explore programming and runtime models which decouple algorithmic complexity, parallelism concerns, and hardware mapping. We build upon these frameworks to exploit domain-specific semantics using high-level transformations and modifications to obtain performance through algorithmic and runtime optimizations.

We first discusses optimizations performed on a computational mechanics solver using a novel coupling technique for multi-time scale methods for discrete finite element domains. We exploit domain semantics using a high-level dynamic runtime scheme to reorder and balance workloads to greatly improve runtime performance. The framework presented automatically chooses a near-optimal coupling solution and runs a work-stealing parallel executor to run effectively on multi-core systems.

In my latter work, I focus on the parallel programming model, Concurrent Collections (CnC), to seamlessly bridge the gap between performance and programmability. Because challenging problems in various domains, not limited to computation mechanics, requires both domain expertise and programming prowess, there is a need for ways to separate those concerns. This thesis describes methods and techniques to obtain scalable performance using CnC programming while limiting the burden of programming. These high level techniques are presented for two high-performance applications corresponding to hydrodynamics and multi-grid solvers.

Download

Share

COinS