SMP Clusters上資源分配之研究

本論文主要研究平行工作（parallel jobs）在SMP叢集上，CPU和網路資源分配之探討。首先在平行工作scheduling演算法方面，如果分配給這些jobs的CPUs的位置，是集中在少數SMPs上，communication的機會較小，因此對網路的負擔也較小，但會增加waiting time；反之，如果分散到許多SMPs上，waiting time會降低，但communication較多，對網路的競爭也較嚴重。本論文以效能較佳的backfilling演算法為實驗對象，我們在模擬器上，利用有公信力的trace，在不同的網路通訊負擔的假設之下，執行各項實驗。本實驗證實：當系統要分配CPUs給job時，可以依照當時的網路的負載程度的輕重，動態調整這些CPUs的集中程度（Tight值），原則是網路的負載程度重時則要求集中，網路的負載程度輕時則放寬集中程度。然後去尋找是否有足夠的CPUs個數且符合Tight值，若有則將這些CPUs分配出去，否則等待其他job(s)釋出CPUs後再行判斷是否可以進行後續的分配工作。

關鍵字

叢集電腦；資源分配；通訊負擔

並列摘要

This paper studies the resource allocation of parallel jobs on SMP clusters. Previous parallel job scheduling algorithms, such as EASY (Extensible Argonne Scheduling sYstem), a backfilling algorithm, focus only on the allocation of CPUs. As communication cost becomes a bottleneck and gradually dominates the performance of program executions on multicomputers, some processor allocation policies suggest that as processors are allocated to a job, the allocated processors must be as continuous as possible. However, some of the mentioned algorithms will lead to another external fragmentation problem. The problem occurs as sufficient processors become available for the requested job; however, these processors are dispersed throughout the system. Thus jobs must wait for some time until there are sufficient contiguous processors. In summary, as adjacent processors are allocated, we may improve the run time of jobs, however, this will increase the wait-time of jobs. In this paper, we suggest that an effective processor allocation policy should well balance communication cost and waiting delay. The principal is simple. When a network is busy or the communication cost is high, the processors must be allocated as continuously as possible. On the other hand, when the communication cost is light‚ to shorten the waiting time‚ the processors should be allocated as soon as possible.

並列關鍵字

cluster of SMPs ； Resource Allocation ； Communication Cost

國際替代計量

SMP Clusters上資源分配之研究

全文下載

主題瀏覽