Font Size: a A A

Parallel Computing Scalability Studies And Applications On The Distributed Memory Environments

Posted on:2001-12-17Degree:DoctorType:Dissertation
Country:ChinaCandidate:J ChenFull Text:PDF
GTID:1118360092998879Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
With the development of supercomputers, more and more attentions have been paid to the research field of the scalability of the combinations of parallel algorithms and parallel computers. To meet these requirements, we developed our study on scaling models, scalability metrics and analysis criterions. We also used the conclusions we got to analyze several applications. The main work we have done is summarized as follows.(1) The scalability studies are based on the studies of parallel computation models. We analyzed some of the preciously presented parallel computation models thoroughly, then we classified the common parallel computation models, and describe their applicability and shortages.(2) We analyzed the relationship between the scalability and the execution time, and the relationship between the scalability and serial performance. The results show that if the execution time or serial performance is emphasized unilaterally, there may be some bad effects on the scalability. We also studied the affection of data partition strategy on the scalability by theoretical analysis and experiments.(3) We put forward a new scaling model named the near-optimal scaling model which is presented from the point of cost-effectiveness for the first time. This model can be used not only to describe the algorithm's scaling capacity, but also to predict how many processors the larger-sized applications required to get the "reasonable" performance based on the performance of smaller parallel system, which means that the execution time is near to the shortest time, and gain the acceptable efficiency at the same time. The new scaling model can help users to optimize their processor configure.(4) We put forward three scalability metrics, named time scalability, efficiency scalability and convergence scalability. Precious scalability metrics can hardly be used to distinguish the scalability differences caused by the algorithm implementation variations. These new scalability metrics overcome this disadvantage, and can be easily measured.(5) We put forward several scalability analysis criterions which can help to determine the reason of why the scalability of the certain practical parallel system is not high enough.(6) By using the above method, we analyzed the scalability of three parallel applications which come from the field of nuclear physics and CFD respectively. These applications are two dimensional electromagnetic plasma with particle-in-cell method program (2D-CIC),tracer interfaces moving grid method program (YGX), three-dimensional supersonic complex base flowfields numerical simulation (TXD). The experiment results can validate the conclusions we got. And we also analyzed the resource requirements of these applications.(7) To the heterogeneous systems, we present an improved performance model, and a static task-partition strategy. Results showed that this strategy can achieve load-balance. We also developed a method to utilize resource effectively when the resources are limited. By applying it to some typical parallel algorithms, we got a conclusion that this method can predict the processor number that the parallel algorithm required to get the shortest execution time on the heterogeneous and resource-limited computing environment.(8) Heterogeneous computing is currently a hot research field, but the previous scalability studies focus on the analysis of the homogenous systems. We presented a new-scaling model for the heterogeneous systems built on the homogenous network. Experiments on some typical algorithms show that it can describe the scaling ability of these heterogeneous systems.(9) We parallelized two serial CFD programs, named the supersonic high-order panel software and the Euler equations numerical simulation program. During the parallelization procedure, we take consideration of the scaling ability of the parallel programs, and explore how to use the scalability studies to guide the parallel programming.
Keywords/Search Tags:scalability, performance evaluation, parallel computing, parallel computation model, heterogeneous computing
PDF Full Text Request
Related items