Analysis And Optimization Of SIMT Thread Scheduling Model

Due to the development of the information world, computer processors need toprocessor more and more information. Because the processor has to have the highability of the data process, traditional micro-processor architectures cannot adapt tothe efficient process of the huge of information. It is a tendency of improving theprocessor performance that the technology develops from the traditionalinstruction-level parallelism to the thread-level parallelism. During the research ofmulti-thread, single instruction multi-thread (SIMT) technology has become one ofthe most concerned architectures. The graphics processing unit (GPU) based on theSIMT technology is a most famous one which has high-efficiency multi-threadscheduling technology and has adapted adequately to data intensive applicationsafter many years development in the field of multi-media.The current SIMT researches orient their work to the whole GPU architecture,which is short of the independent study of SIMT scheduling model extracted fromGPU. This article analyzes the SIMT scheduling technology in the GPU andextracts the SIMT scheduling model from the GPU architecture in order to lay theresearch foundation for the application and development of the SIMT schedulingtechnology. What’s more, this article carries on the detailed analysis of SIMTscheduling model, which finds that the organization management of threads has agreat influence on the performance of SIMT scheduling model. At the same time,the request of users who use the SIMT scheduling technique is high, which meansthat the technology level of researchers on the hardware and software is high. Theworkloads must fully matched SIMT hardware features, at the same time change theSIMT scheduling architecture must adapt itself to the workloads’ characteristics,because different workloads’ characteristics on SIMT hardware structure will havegreat differences.Through the analysis of the SIMT scheduling model, this article finds that theimplementation capacity of single thread existing in SIMT scheduling technology islow and that it must rely on a large number of threads to hide poor single-threadperformance disadvantage. So the dependence of SIMT scheduling techniques onthe number of threads is strong, which is the urgent need to solve the technicalproblems in the development of it. This article finds that characteristics of aninnovative explicit data graph execution (EDGE) architecture are similar with thescheduling technology characteristics of SIMT through researching differentcomputer architectures. This paper finds that SIMT scheduling techniques can not only dig out multithreading features in EDGE potential in it, but also can use theexecution efficiency of a single thread in EDGE to correct the single threadperformance disadvantage in the SIMT scheduling technique and to optimize theSIMT scheduling model. In this paper, the SIMT scheduling technology isimplemented on M5-EDGE simulator developed by the lab. This article verifies thecorrectness of SIMT scheduling function in the M5-EDGE simulator by a series oftest program verification and analyzes the performance of SIMT schedulingtechnique implemented in the M5-EDGE simulator. It finds that single threadhyperblock parallelism in the EDGE distributed structure can be well combinedwith multithreading parallelism of SIMT scheduling technology, which give fullplay to the advantage of the EDGE and the SIMT architecture.