集群计算环境下高性能通信库HPCL的设计与实现

Design and Implementation of High Performance Communication Library on Cluster

论述了一个高性能通信库系统—HPCL的研究与实现.HPCL的技术特点包括支持多路通道、允许多个进程同时通信、支持可靠的消息传递、支持通道上下文切换、允许一个通道被多个进程按照分时的方式来使用、支持多点传送,具有低延迟,高吞吐量等.这些特点的实现技术包括改进的ACK/NACK流控制算法,基于快速Ethernet网络接口的及时发送(ImmediateSending)技术、中断回收(InterruptReaping)技术.使用典型Benchmark对HPCL和TCP/IP的通信性能、应用性能进行对比表明,HPCL达到了理想的吞吐率和时间延迟.

We have developed a high performance communication library called HPCL. HPCL supports multiple communication channels ,allowing multiple processes communicate simultaneously ; Guaranteeing message delivery ; channel context switching, allowing a channel to be used by multiple processes in a time-sharing fashion;Multicast assistance, provided with low latency. high throughput and other excellent characteristics. The modified ACK/NACK flow control algorithm has been developed to realize these features. In order to implement HPCL based on Ethernet network interface,in this paper,authors also use several techniques such as immediate sending,Interrupt reaping and so on. We also compare the communication and application performance of HPCL and TCP/IP,using the same benchmarks. The results show that the HPCL can achieve best performance in throughput ,latency and application.