摘要
InfiniBand是目前HPC系统互连的主流网络之一,其提供的可靠连接传输服务因为支持RDMA、原子操作等功能而被广泛应用于MPI等并行应用编程模型。但是支撑可靠连接所需的消息队列及缓冲区开销往往会随着并行规模的扩大而急剧增加,从而制约了应用规模的扩大。为了解决这种内存开销带来的消息可扩展性问题,先从In-finiBand传输优化方面介绍了共享接收队列和扩展可靠连接技术,然后基于并行通信模型提出了分组连接技术。通过这些技术可以将节点内存开销减少2个数量级,并且开销不会随并行规模的扩大而明显增加。
InfiniBand is one of the most promising network interconnecting technologies in HPC. Its reliable connection services support RDMA, atomic operations, ete, and are widely employed in MPI and other similar parallel programming models. However,the cost of message queues and buffers for reliable connections rises dramatically when the scale of parallelism increases. As a result, it becomes the bottleneck for InfiniBand in large scale applications. To solve the pro- blem, this paper provided shared receiving queue and extended reliable connection, then brought forward the group con- nection technology based parallel communication mode. On hand, the memory cost in the computing node is cut down for at least 100 times smaller. And the most amazing thing is that the cost of memory will be relatively constant when the scale of parallelism is upgraded.
出处
《计算机科学》
CSCD
北大核心
2013年第3期104-106,120,共4页
Computer Science
