(A) link-disjoint submesh for processor allocation in mesh computers

Abstract

A processor allocation scheme for mesh computers greatly affects their system utilization. The performance of an allocation scheme is largely dependent on its ability to detect available submeshes. We propose a new type of submeshes, called link-disjoint(LD)submeshes ,for processor allocation in mesh computers. This type of submeshes increases the submesh recognition capability of an allocation scheme. The adaptive scan(AS) strategy [Ding93], [Sharma9], the quick allocation(QA) strategy [Yoo95] can find all the available contiguous submeshes whose processors are constrained to be physically adjacent. In [Ding93],they show that the processor utilization using the AS strategy, is about 45% in 64 x 64 mesh systems when the probability of small-sized submesh request is greater than that of large-sized submesh request. It is apparent that if more submeshes are available, better job response time and processor utilization can be achieved. In the previous strategies, the constraint that all processors of a submesh must be adjacent causes the external fragmentation of processors even with idle processors. A link-disjoint submesh consists of contiguous and/or non-contiguous processors, but the links between non-contiguous processors are assigned to have no communication conflict with any other submeshes. When wormhole routing or circuit switching is used, the performance degradation caused by non-contiguous processor allocation is almost minor. We devise the LD algorithm to find out all possible submeshes including link-disjoint submeshes. Through simulation, the performance of our scheme is measured and compared with the previous methods. It is shown through simulation that the link-disjoint submeshes increase the performance of an allocation scheme. In our simulation, the system utilization of our scheme is about 10% more than the previous methods. Moreover, this link-disjoint concept can be easily applied to higher dimensional mesh computers, such as three d...