Expedite : An Operating System Extension to Support Low -latency Communication in Non- dedicated Clusters
Expedite : An Operating System Extension to Support Low -latency Communication in Non- dedicated Clusters
复制标题
Expedite：支持非专用集群中低延迟通信的操作系统扩展
DOI：
--
复制
发表时间：
2004
影响因子：
0.6
通讯作者：
Chikayama Takashi
中科院分区：
文献类型：
--
作者：
Saito Hideo;T. Kenjiro;Chikayama Takashi
我们提出并评估了一种简单的操作系统扩展，允许并行应用程序在非专用集群中以低延迟进行通信。我们扩展了UNIXAPI，允许进程设置套接字的O加速标志，并修改Linux调度器，以便当数据到达设置了O加速标志的套接字时，为进程提供较大的优先级提升。通过只将提升给予大多数时间阻塞的进程，并通过使优先级提升快速衰减来维护公平性。我们描述了一个利用此功能的MPI实现。实验表明，即使当许多高优先级进程在后台运行时，使用我们的工具也可以减少集合操作(如MPI散布())的延迟。
We propose and evaluate Expedite, a simple operating system extension that allows parallel applications to communicate with low latency in non-dedicated clusters. We extend the UNIX API to allow a process to set the O EXPEDITE flag of a socket, and revise the Linux scheduler so that a process is given a large priority boost when data arrives on a socket with the O EXPEDITE flag set. Fairness is maintained by giving the boost only to processes that block most of the time, and by making the priority boost decay quickly. We describe an MPI implementation that takes advantage of this facility. Experiments show that the latency of collective operations such as MPI Scatter() can be reduced with our facility, even when many high-priority processes are running in the background.