University of Houston
Department of Computer Science

In partial fulfillment of the Requirements for the Degree of
Master of Science


Mohamad Chaarawi
will defend his thesis

Optimizations of Group and Communicator Operations in Open-MPI

Abstract
MPI is the de-facto standard for message passing in scientific applications. Open MPI is a new MPI2 implementation, combining members from industry and academics with many decades of combined experience in developing message passing libraries. The current implementation of the process group functionality within Open MPI leaves however room for many optimizations.
The goal of this study is to optimize process group management functions of the library in terms of memory consumption and communication costs. The first goal is achieved by providing alternative implementations for MPI groups by introducing sparse storage formats for processes of large groups in case the difference between two large groups can be described in a compact way.
The second part of the study aims to optimize the algorithm used for obtaining a context ID for a newly created communicator. The current algorithm is considered expensive since it requires multiple collective communication operations. The new algorithms implemented will aim to reduce the number of collective operations and hence reduce the latency for communicator creation.
Both improvements are evaluated on a 24 node dual-core AMD Opteron cluster, using both the InfiniBand and the Gigabit Ethernet network interconnects.

Date: Tuesday, November 21, 2006
Time: 1:00 PM
Place: 362-PGH

Faculty, students, and the general public are invited.
Thesis Advisor: Dr. Edgar Gabriel