Grid computing involves employing multiple machines to solve complex computing problems. For my project I have simulated a grid computing system based on a cluster of machines located at the University of Calgary. Data was collected from this system over a period of seven days. I have examined scheduling of jobs to nodes in the computing cluster. In particular I have simulated single site, centralized scheduling algorithms. This means that I assumed all jobs enter one queue and are then assigned to machines based on scheduling criteria.
Through my analysis of the collected data I was able to conclude that the arrival rate of jobs entering the grid computing system followed a Poisson distribution with λ = 3:125 jobs per half hour. I found that the running time of processes followed a negative exponential distribution with a mean of 496:3 minutes. Using these distributions I simulated this system using GPSS-H. From my results I can conclude that in a system where the majority of jobs require less than 4 processors it is reasonable to allow these jobs to contend for 4-processor machines. Also, using a most nodes available selection strategy versus a random selection strategy does not seem to enhance performance significantly.
Source: University of Toronto
Authors: S.Panzieri | F.Pascucci | G.Ulivi