Simulation for bulk synchronous parallel superstep task assignment in desktop grids characterised by gaussian parameter distributions

摘要

Modern load distribution schemes for parallel task assignment often refer to parallel computers with shared memory or dedicated clusters with local address spaces. For parallel computing performed on dedicated clusters, the communication topology between processors can be considered static. On dealing with the execution of parallel programs on desktop grids, the previous assumption will lead to inexact parameter determination for network cost due to the actual network topology being determined not only by the hardware characteristics but also by the instantaneous bandwidth availability for the processes to communicate. The dynamic bandwidth depends not only on location, time and date, but also on the habits and behaviour of the individual users of the computers that conform the grid environment and will lead to an inconstant network topology. This paper presents a novel simulator design which takes these factors into consideration for the BSP parallel computer model. The simulator is used as a performance evaluation tool for several case studies to point out variations in performance obtained from different task assignment strategies as the size of the desktop grid increases and the number of parallel threads changes.