CPU-GPU异构系统下的片上网络仲裁机制研究

摘要

在CPU-GPU异构系统架构中,由于GPU程序的多线程特点,大多数GPU程序会垄断系统中的共享资源,例如片上网络.这将给CPU程序的性能造成很大的损失.我们发现有一些GPU程序性能对网络延迟表现为不敏感.当CPU、GPU程序同时竞争片上网络时,降低这类对网络延迟不敏感的GPU程序的优先级,提高CPU程序的优先级,结果会对GPU程序的性能会造成很小的损失,同时可以大幅提高CPU程序的性能.因此,我们可以通过适时提高CPU程序的网络使用优先级来优化CPU程序的性能.当前通用的循环调度片上网络仲裁机制并没有考虑CPU、GPU程序不同的特性,系统性能存在提升的空间.基于此,我们提出一种新的基于网络延迟敏感度的仲裁机制.实验结果显示,相对于循环调度片上网络仲裁机制系统而言,基于延迟敏感度仲裁机制系统,CPU程序性能有16.6％的提高,而GPU程序性能发生3.65％的下降.%In CPU-GPU heterogeneous architecture,most GPU applications tend to monopoly sharing resource for example on-chip network,due to GPU applications' high thread level parallelism.CPU applications' performance suffers great loss.We investigate that some GPU applications are insensitive to network's latency,decreasing these GPU applications' priority on using network will cause little influence to their performance,at the same time CPU applications can get benefit a lot from it.We can optimize CPU performance by scheduling the priority,when CPU application and GPU application contend for the network.While current common network round robin arbitration mechanism does not consider the different characteristic of CPU and GPU application.We propose a network latency sensitivity based arbitration mechanism.The experiment result is CPU performance gets 16.6％ improvement,while GPU performance gets 3.65％ loss over baseline:round robin arbitration mechanism.