In Partial Fulfillment of the Requirements for the Degree of
Master of Science
Will defend his thesis
In this work, we simulate the performance of parallel applications on desktop grids. We provide a framework to effectively estimate the effects of varying bandwidth, latency and node failure on the performance of parallel applications. The framework is used to estimate the usage potential of desktop grids to run parallel applications and is instrumental in knowing what kind of applications perform reasonably well in such computing environments. The experiments have been conducted using NAS parallel benchmarks, by collecting the trace of their execution on Shark cluster, which is a collection of 29 nodes connected through a 96 port 4xInfiniBand switch, and simulating their performance using Dimemas, which is a performance simulation tool. The types of network simulated include actual bandwidth and latency measurements taken in different labs at University of Houston. The simulation results for actual bandwidth and latency measurements show us that on average, the applications perform 2x, 4x and 6x slower in the minimum, average and maximum case when deployed on desktop grids consisting of machines in a single lab compared to Shark cluster, which is a dedicated parallel machine. Parallel benchmarks perform faster by 2x, 3x and 5x in the minimum, average and maximum case on machines in a single lab environment compared to being deployed on machines in distributed labs connected through WAN/intranet. The results show that desktop grid is a suitable platform to deploy and run parallel applications by incurring minimal performance degradation.