面向多任务环境的MCM-GPU关键技术研究

多芯片模块GPU即MCM-GPU和高效支持多任务成为GPU发展的必然趋势。在支持多任务的MCM-GPU中，运行的多个任务具有不同的特征，使MCM-GPU面临着共享资源优化管理和冲突处理等诸多挑战。为了缓解这些挑战，需要在分析程序多样性的基础上优化设计MCM-GPU架构。本课题主要研究高效支持多任务环境的MCM-GPU关键技术，包括（1）计算资源流处理器分配管理，在多任务运行时根据每个任务的特征动态地为其分配合适数量的流处理器；（2）通信资源片上网络共享优化，针对MCM-GPU新型的片上网络进行任务调度和片上网络资源管理；（3）存储资源末级高速缓存高效设计，满足同时运行的多个任务对MCM-GPU中末级高速缓存的不同需求；（4）共享资源访问冲突处理，针对MCM-GPU的特点构建低开销的混合减速预测模型准确的预测单个任务由于共享资源访问冲突导致的性能损失。

Multi-chip-module GPU, i.e., MCM-GPU, and multitasking GPU are becoming more and more important. In MCM-GPU with multitasking support, the co-executing applications have various characteristics that poses challenges to the shared resource management and contention handling schemes in multitasking GPUs. To solve these challenges, we need to optimize the multitasking GPU architecture based on analyzing the application characteristics. This research proposal is about the key techniques to effectively support multitasking on GPUs, including (1) the online SM allocation scheme which finds the effective SM partition between different applications during the execution, (2) NoC-aware task scheduling and resource management which targets the new scalable MCM-GPU NoC, (3) effective LLC design of MCM-GPUs to meet the different LLC demands of co-executing applications and (4) a low-cost hybrid slowdown prediction model which accurately predicts the performance of a single application under the resource contention based on the MCM-GPU feature.