next up previous
Next: Test Problems and Up: Computational Environments Previous: Performance Characterizations of

4.2 Performance Evaluation of the Parallel Implementation

  To evaluate the speed of execution for the parallel CSI-MSVD algorithm, the elapsed wall-clock time (in seconds) was monitored. Although the time taken to start up all PVM processes and to print the final results were not included, the elapsed time for loading the input matrix into memory was taken into account. Timings were obtained on a network of 24 machines on a LAN (10Mbps Ethernet or 100Mbps FastEthernet) isolated by a bridge to avoid interference from external traffic on the Internet, and thus provide an estimate of the best-case performance. Each machine on the LAN was a Sun SPARCstation 5 Model 70 workstation with a clock speed of 70 MHz and 32 MB memory (later upgraded to 96 MB memory).

In this work, speedup is used to estimate the scalability of the algorithm with a variation in problem size, so that for a given class of problems, recommendations can be made for the optimal choice of virtual machine configuration. Speedup is defined as


where is taken to be the wall-clock time for CSI-MSVD to execute on one processor without any redundant synchronization/communication operations and is the wall-clock time with p processors. Although this definition of does not necessarily reflect the wall-clock time of the best single-processor algorithm, the objective in computing S through Equation (38) is to obtain a measure of the speedup over single-processor time, i.e. the scalability of the system.

Michael W. Berry (
Sun May 19 11:34:27 EDT 1996