Department of Computer Science at UH

University of Houston

Department of Computer Science

In Partial Fulfillment of the Requirements for the Degree of
Master of Science

Spoorthy Mareddy

Will defend her thesis

Evaluating and Optimizing the Performance of
Hierarchical Collective Operations


The performance of collective operations is crucial for high performance computing. As the optimal performance of a collective operation for a given system depends on many factors, Open MPI offers several components for MPI collective communications and also provides the selection logic for choosing the best suitable component for the given system. The hierarchical collective component is used when a communicator spans over multiple nodes and is designed to exploit the hierarchy in networks to optimize collective operations. But, the overhead imposed by hierarchy detection algorithms is significant. The goal of this thesis is to improve the performance of hierarchical collective component by reducing the time spent in hierarchy detection. Two different hierarchy detection algorithms, namely, two-level hierarchy detection algorithm and prefetch method, have been designed and implemented. The algorithms are evaluated as part of this thesis and performance improvement is seen in many cases.

The second part of this thesis deals with determining optimal parameters for collective operations within a given execution environment. The Open Tool for Parameter Optimization (OTPO) is a tool used for tuning runtime parameters of Open MPI. Within this thesis, OTPO is extended to explore the parameter space of hierarchical collectives using the popular SKaMPI benchmark.

Date: Tuesday, April 28, 2009
Time: 10:00 AM
Place: 550-PGH
Faculty, students, and the general public are invited.
Advisor: Prof. Edgar Gabriel