通用圖形處理器程序在圖形處理器叢集上之動態遷移

Abstract

Unlike classic video cards, modern Graphics Processing Unit (GPU) has been widely used for general-purpose computation acceleration. GPU clusters have become very popular in the field of High Performance Computing (HPC), three of the top five super-computers utilize GPUs. Migration is a critical capability in such large-scale GPU cluster system environment because it supports resource management, load balancing, system maintenance and fault tolerance. In this thesis, we propose a mechanism for GPGPU applications. We have two contributions. The first contribution is we support preempt-like functionality in GPUs. When migration is triggered, CPU can pro-actively terminate GPU compu- tation instead of waiting task complete, therefore the migration can start earlier. The second contribution is we hide data transfer overhead by overlapping compu- tation with transfers. We send back the part of completion data while GPU is computing. When migration is triggered, we pre-copy the part of completion data to overlap computation with transfers. In the experimental results, we demonstrate that our proposed mechanism quickly releases GPU resource when migration is triggered and also decreases the migration overhead resulted from transfers.