文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > Building a High-Performance Collective Communication Library
侵权投诉

Building a High-Performance Collective Communication Library

In this paper, we report on a project to develop a unified approach for building a library of collective communication operations that performs well on a cross-section of problems encountered in real applications. The target architecture is a two-dimension

Building a High-Performance Collective Communication Library Mike Barnett Department of Computer Science University of Idaho Moscow, Idaho 83844-1010 Lance Shuler Parallel Computing Sciences Department, 1424 Sandia National Laboratory Albuquerque, New Mexico 87185-1109 Satya Gupta and David G. Payne Supercomputer Systems Division Intel Corporation 15201 N.W. Greenbrier Pkwy Beaverton, Oregon 97006 Robert van de Geijn Department of Computer Sciences The University of Texas at Austin Austin, Texas 78712{1188

Jerrell Watts Computer Science Department California Institute of Technology Pasadena, California 91125

Abstract In this paper, we report on a project to develop a uni ed approach for building a library of collective communication operations that performs well on a cross-section of problems encountered in real applications. The target architecture is a two-dimensional mesh with worm-hole routing, but the techniques are more general. The approach di ers from traditional library implementations in that we address the need for implementations that perform well for various sized vectors and grid dimensions, including non-power-oftwo grids. We show how a general approach to hybrid algorithms yields performance across the entire range of vector lengths. Moreover, many scalable implementations of application libraries require collective communication within groups of nodes. Our approach yields the same kind of performance for group collective communication. Results from the Intel Paragon system are included. To obtain this library for Intel systems contact intercom@cs.utexas.edu.

monly used collective communication algorithms. It is this emphasis on a high-performance implementation that sets it aside from the MPI e ort 16], which tries to standardize the interface to communication libraries. Indeed, we expect the fruits of our e orts to be incorporated into implementations of the MPI standard. The following collective communication operations have been identi ed as being useful in many applications: broadcast, scatter, gather, collect and global combine. Typical approaches to implementing such collective communication algorithms are limited to considering the case of short vectors, which are treated with one technique, or to considering the case of long vectors, for which very di erent techniques are appropriate. For a general purpose library, it is crucial that an implementation performs well for all vector lengths. In an earlier paper 1], we have presented a general approach to building collective communication libraries. In this paper, we give a further description of this approach, with emphasis on an explanation of hybrids.

1 Introduction The Interprocessor Collective Communication (InterCom) Project is a comprehensive study of techniques for a high-performance implementation of com-

2 Target Architectures Our current implementation assumes a twodimensional physical mesh of processing nodes, with

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email