基于Infiniband网络的消息传输技术研究

Infiniband的Send/Receive通信语义中,接收请求都必须先于发送请求被提交,错误的操作则会导致数据传输的延迟或者失败,支持数据传输的队列和缓冲区的开销会随着并行规模的增大而急剧增加,严重制约了并行规模的扩大.为了解决Infiniband工作请求提交顺序和内存开销的问题,进一步提高Infiniband的传输效率,本文先提出了一种程序控制流的方法保证请求被正确的执行,然后提出一种共享内存缓冲区技术减少内存的开销,最后通过循环使用多个队列的办法提高Infiniband的传输效率.在并行绘制的集群系统中,本文对改进后的Infiniband网络的传输性能进行了测试和分析.

Research of message passing technology based on infiniband architecture

During every communication by Infiniband Send/Receive, receive request must be submitted prior to send request, wrong operation can cause delay or failure of the data transmission. The overhead of queues and buffers which support the data transfer rapid increase with the increase in the parallel scale, restricted the parallel expansion of the scale severely. In order to solve problem of order that Infiniband work request submitted and memory overhead, further improve the transmission efficiency of Infiniband. First, this paper proposed a program control flow method to ensure proper completion of the every work request, and then proposed techniques of a shared memory buffers to reduce the memory overhead, at last, proposed the idea of cycling using multiple queues to improve the transmission efficiency of Infiniband. In this paper, the transmission performance of the improved Infiniband network were tested and analyzed.