Date of Award

Fall 2013

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering

First Advisor

Samuel P. Midkiff

Committee Chair

Samuel P. Midkiff

Committee Member 1

Mithuna S. Thottethodi

Committee Member 2

Rudolf Eigenmann

Committee Member 3

Vijay S. Pai

Abstract

Bugs in sequential programs cost the software industry billions of dollars in lost productivity each year. Even if simple parallel programming models are created, they will not reduce the level of sequential bugs in programs below that of sequential programs. It can be argued that the complexity of current parallel programming models may increase the number of sequential bugs in parallel programs because they distract the programmer from the core logic of the program.

Tools exist that identify statements related to sequential bugs and allow those bugs to be more quickly located and fixed. Their use in parallel programs will continue to be useful. Many of these debugging tools require runtime monitoring of program points of interest in a program and the overhead of this monitoring is usually very high.

We propose Ant, a framework that increases the efficiency of sequential debugging techniques when used with parallel programs. The Ant framework takes two different strategies depending on whether the program to be debugged is a distributed memory program or shared memory program. For MPI programs, the Ant compiler analyzes the program and identifies two different types of code regions: those that all processes execute and regions that only part of the processes execute. For shared memory Pthreads programs, Ant uses a combination of static and dynamic analyses to determine similar parts of the program executing in parallel and the number of threads executing those parts of the program. The programs are instrumented with calls to Ant runtime libraries and debugging libraries based on the Ant compiler's static analysis results. Relative to a naive port of a debugging tool (C-DIDUCE, in our cases), Ant's technique, by exploiting the application's parallelism, reduces the monitoring overhead by up to 15.85 times (and on average 9.23 times) for MPI programs executing with 32 processes and up to 18.14 times (and on average 8.73 times) for Pthreads programs executing with 8 threads, while maintaining high accuracy.

Share

COinS