ハイブリッドTCと非同期MPI-SMPクラスタにおける高性能プログラミングモデル

Nowadays, cluster of SMPs has been becoming the main architecture in the super computing world. Clusters of the latest Intel multi-core processors are among examples. With the arising popularity of SMP clusters, it is extremely important to develop an effective SMP-aware parallel programming model. The model should supply users with high-performance, portability, and simplicity. That is, while providing the performance better than non SMP-aware models, it has to be available to ａ wide range of underlying hardwares and operating systems, and be simple enough in writing and debugging for common parallel programming users. For those reasons, this study\u27s object is to develop effective SMP-aware programming models for SMP clusters. Before this study, flat MPI had been widely accepted as the most effective programming model in spite of several SMP-aware attempts. However, finding out that flat MPI wastes many computational and communicational resources, we propose ａ node level communication computation overlap that allows the clusters to use their resources more effectively and thereby, improving overall performance. With the overlap, we can eliminate bottlenecks on the memory bus and the network connection. On an SMP node, several processors share different kinds of resources such as memory bandwidth and one or more network connections. The de-facto flat MPI model treats all processors of an SMP equally and in most the cases forces them work synchronously. This may result on network saturation with a free memory bus bandwidth during system communication time, or memory bus saturation with a free network bandwidth during system computation time and thereby, slowdown the overall performance. Within the proposed overlap, when some processors perform computation, the remainders do communication. As ａ result, we can make use of a11 kinds of shared resources at any point in times and avoid the bottlenecks. We then realized the overlap idea by implementing Hybrid TC and Asynchronous MPI. We also provide a formula to evaluate benefit the new models can achieve without actual implementation and prove its correctness through several experimental results. Hybrid TC showed good performance. In the model, an SMP node runs only ａ single MPI process with several OpenMP threads, each on ａ processor. A single thread performs all of the communication tasks of ａ node. The remainders get necessary data through shared variables located on the shared memory area. We also solved obstacles on the implementation process. That is, ａ new task schedule in which main communication and computation tasks have no data dependency and a solution for the grain size problem. While applying the model in solving the High Performance Linpack Benchmark (HPL) problem. which is ａ global meter to evaluate actual performance and rank the most powerful systems in the world, the performance was up to 57% better than that of the flat MPI. We then proposed the asynchronous MPI model that provides equivalent performance as hybrid TC. In the model, each processor runs an MPI process: processors of an SMP node run in difference phases, such that while some are performing computation, the remainders are performing communication tasks. Its task schedule becomes much simpler than that of hybrid TC. As ａ result, we can form the overlap and get a hybrid TC equivalent performance while reducing programming efforts significantly. Both these two implementations are quite effective and simple for implementa- tion. While asynchronous MPI does not require OpenMP, hybrid TC solutions are more flexible with varying number of processors per node. Our main contributions in this study are listed as follows: 1.The node level over lap : by proposing the overlap, we determine the essential principle to acquire high performance onSMP clusters. 2.HYbrid, TC and asynchronous MPI: the two implementations of the node level overlap. The two models achieve the improvement of up to 57% of performance in comparison with the de-facto flat MPT model. 3.The simplification in implementation. Our study develops techniques and algorithms that allow common parallel programmers to build effective overlapping solution from an existing flat MPI code with minimum programming efforts. 4.A formula for evaluating over lap performance: using this formula, user can forecast the improvement that overlap solutions can achieve beforehand, and decide whether it is worth to reconstruct the existing MPI code.電気通信大学200