摘要: :针对集群计算系统不同的底层通信库,高效率地实现MPI函数是提高集群应用系统效率的有效途径。MPI_HPCL(Massage Passing Interface Based on HPCL_USTB)就是在集群系统底层通信库HPCL_USTB(High Performance Communication Lib of USTB)的基础上实现的具有零拷贝支持的MPI系统。它能够有效地管理被锁定的存储空间,避免死锁现象,采用对发送和接收的锁定存储进行分开控制和延迟队列的策略,保证了零拷贝消息传送的高效率和安全可靠。
关键词:
集群计算,
MPI_HPCL,
HPCL_USTB,
并行计算
Abstract: It is an effective way to advance the cluster system efficiency that implements the MPI functions efficiently based on those various cluster low-level communication libraries. MPI_HPCL(Massage Passing Interface Based HPCL_USTB) is just an implementation of zero-copy MPI on the top of a cluster low-level communication library named HPCL_USTB(High Performance Communication Lib of USTB). It is capable of managing pin-down memory and avoiding deadlock. MPI_HPCL realizes the characteristics of good efficiency, safety and reliability by adopting some methods, such as separating the control of the send/receive pin-down memory and using the delayed queue.
Key words:
Cluster computing,
MPI_HPCL,
HPCL_USTB,
Parallel computing
尹怡欣;邱 博;胡长军;徐正光;王 珏. 集群计算通信库HPCL_USTB的零拷贝MPI系统[J]. 计算机工程, 2006, 32(23): 58-60.
YIN Yixin; QIU Bo; HU Changjun; XU Zhengguang; WANG Jue. Zero-copy MPI Based on Cluster Computing Communication Library HPCL_USTB[J]. Computer Engineering, 2006, 32(23): 58-60.