The Influence of Communication on the Performance of Co-allocation

Abstract

In systems consisting of multiple clusters of processors interconnected by relatively slow connections such as our Distributed ASCI 1 Supercomputer (DAS), jobs may request coallocation, i.e., the simultaneous allocation of processors in different clusters. The performance of co-allocation may be severely impacted by the slow intercluster connections, and by the types of job requests. We distinguish different job request types ranging from ordered requests that specify the numbers of processors needed in each of the clusters, to flexible requests that only specify a total. We simulate multicluster systems with the FCFS policy-and with two policies for placing a flexible request, one tries to balance cluster loads and one tries to fill clusters completely-to determine the response times under workloads consisting of a single or of different request types for different communication speeds across the intercluster connections. In addition to a synthetic workload, we also consider a workload derived from measurements of a real application on the DAS. We find that the communication speed difference has a severe impact on response times, that a relatively small amount of capacity is lost due to communication, and that for a mix of request types, the performance is determined not only by the separate behaviours of the different types of requests, but also by the way in which they interact.

Parallel Job Scheduling — A Status Report

“…In this situation, the network performance varies greatly from communication within a parallel computer to communication between two parallel computers. Some models have been derived to evaluate the performance of so called multi site computing [24,4,9,11,13]. However in practice, such an approach has not been implemented with the possible exception of the preplanned combination of a few specific parallel computers for a specific purpose.…”

Multisite co-allocation algorithms for computational grid

“…Finally, they presented results showing that the use of partitioned configuration did not necessarily imply a performance drawback [21] . Another important research about co-allocation or multisite site scheduling is presented in [22][23][24][25][26][27][28][29][30][31] . Bucur and Epema [22][23][24] assess the influence on the mean response time of the job structure and size, the sizes of the clusters in the system, the ratio of the speeds of local and wide-area communications, and of the presence of a single or of multiple queues in the system.…”

“…Another important research about co-allocation or multisite site scheduling is presented in [22][23][24][25][26][27][28][29][30][31] . Bucur and Epema [22][23][24] assess the influence on the mean response time of the job structure and size, the sizes of the clusters in the system, the ratio of the speeds of local and wide-area communications, and of the presence of a single or of multiple queues in the system. Also, they evaluated different scheduling policies for co-allocation, with unordered requests, in multicluster systems with space sharing for rigid multicomponent jobs [25][26][27][28] .…”

“…For example, many singlesite scheduling algorithms in section 2.2.1 are extended from the traditional heterogeneous computing systems in the same domain, which neglects the possibility of coallocation or multisite computing jobs across the boundary of sites [12][13][14][15][16][17] . Although many papers in multisite job scheduling have addressed the coallocation problem indeed, all the sites from different domains are assumed to be homogenous without taking the heterogeneity in a real-world grid scenario into consideration [18][19][20][21][22][23][24][25][26][27][28][29][30][31] . Moreover, less attention is paid on the adaptability and some algorithms lack scalability [16,17] .…”

Towards Feasible and Effective Load Sharing in a Heterogeneous Computational Grid

“…The support of multi-site parallel execution [12,13,14,15,16] on a computational grid has been examined in previous works, concerning the execution of a job in parallel at different sites. Under the condition of a limited communication overhead, the results from our previous work [10] and from [1,3,4] all showed that multi-site parallel execution can improve the overall average response time.…”