The Implementation and Evaluation of Parallel Algorithms on C.mmp.
Abstract
This dissertation demonstrates the implementation and evaluation of parallel algorithms on C.mmp, a multiprocessor computer system. Initial attempts to demonstrate the performance of a simple parellel algorithm yielded unexpectedly large performance degradations from the theoretical calculations. This unexpected result spawned a study of the C.mmp system to discover and measure the major sources that perturbed the performance of the parallel algorithm. The performance study was conducted at several levels: basic hardware measurements; runtime performance of Hydra, C.mmp's operating system; and overall performance of a particular application: a parallel rootfinding algorithm. The results of this study identified six major sources of performance perturbation. The six sources, in order of importance, were: variations in the compute time to perform the repetitive calculation; memory contention caused by finite memory bandwidth; the operating system's scheduling processes can become a bottleneck; variations in the individual processor speeds; interrupts associated with I/O device service routines; and variations in the individual memory bank speeds.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 01, 1978
- Accession Number
- ADA063616
Entities
People
- Peter N. Oleinick
Organizations
- Carnegie Mellon University