Experimental Study on Time and Space Sharing on the PowerXplorer

Scheduling algorithms in parallel computers fall into two basic categories: time and space sharing algorithms. Space-sharing based processor allocation algorithms can be contiguous or non-contiguous. Studies show that non-contiguous allocation is superior due to decrease in fragmentation. Other studies have reported that executing jobs on fewer processors (folding) can improve the performance of contiguous and non-contiguous allocation. However, the problem with folding is that it is not always applicable because of parallel programming languages and parallel operating systems limitations. Most of previous studies used simulation. Our study is an experimental one for studying time and space sharing on a real parallel machine (the PowerXplorer), with eight processors arranged as a two-dimensional mesh. A set of five scientific applications with differing communication characteristics were implemented and executed using time and space sharing. The observed execution times were used to study and compare time-sharing and contiguous and non-contiguous space sharing with and without folding. Our study showed that time-sharing gave comparable results to space sharing allocation. Further, non-contiguous allocation gave better results than contiguous allocation when folding is not supported. However, when folding is supported contiguous allocation gave the best mean turnaround times.