Automatic Task Assignment System On GPU Cluster

Due to low cost, increasing computing power and friendly programming environment, GPUs have became the mainstream accelerator in modern supercomputers and HPC clusters. The using of GPU accelerator making GPU cluster heterogeneous on the internal resources in system level. computing resource of single node varies greatly. There are not only cpu node which include single cpu or multi cpus, but also include cpu-gpu node which contains single-or multi-GPUs. GPUs are computing resource which supply large-scale data parallel computing. Accordingly, GPU cluster has show its hierarchical parallel computing power, not only include single program multiple data(SPMD) and multiple program multiple data(MPMD) computing, but also support SPMD which is finer granularity for large-scale data and single instruction multiple data(SIMD) computing.MPI+CUDA are the mainstream programming model of currently GPU Cluster architecture, However, Using such a low level programming model, programmers requiring detailed knowledge of the underlying architecture, which exert a heavy burden. Besides, the program will be less portability and inefficient.GPU cluster application are consist of variety tasks, including task which only need cpu resouce and task which need cpu and gpu collaborate to complete the computing work. The simple and random task assignment mechanism of mpi can not finish the mapping of custer tasks and system resources automatically and efficiently. Consequently, programmer must allocate custer tasks carefully which means that programmer requiring detailed knowledge of the underlying architecture, such as, system resource and communication topology. Therefore, the design of cluster application is closely related with the execution platform, which excerts a heavy burden to programmer. Besides, program are of poor portability, when the underlying system architecture changes, it must be reconstructed to adapt to the new system.In this paper, we propose StreamMAP, an automatic task assignment system on GPU Clusters. The main contributions of StreamMAP are as follows:(1)We provide powerful, yet concise language extension suitable to describe the compute resource demands of Cluster tasks.(2)We have developed a run time system to maintain resource information, and to supply an automatic task assignment for GPU Cluster. Experiments show that StreamMAP provide programmability, portability and scalability for GPU Cluster application.