Accelerating MPI _ Reduce with FPGAs in the Network Extended Abstract

Accelerating MPI _ Reduce with FPGAs in the Network Extended Abstract
复制标题

加速 MPI _Reduce 在网络扩展摘要中使用 FPGA

DOI:
--
复制
发表时间:
2017
期刊:
影响因子:
--
通讯作者:
A. Skjellum
A. Skjellum
中科院分区:
--
文献类型:
--
作者:
Joshua Stern;Q. Xiong;A. Skjellum

文献摘要

被引文献

相似文献

MPI集体操作通常是HPC应用程序中的性能杀手,特别是那些需要大量通信和计算的应用程序,如MPI_Reduce。使用FPGA,它提供了通信和计算耦合的能力,我们设计了一个硬件加速器,在网络中实现MPI_Reduce。通过我们的设计,初步的结果表明,我们可以实现2.5倍的加速长消息和10倍的加速短消息比传统的集群。
MPI collective operations can often be performance killers in HPC applications, especially ones that require both heavy communication and computation such as MPI_Reduce. Using FPGAs, which provide the ability to couple communication and computation, we have designed a hardware accelerator to implement MPI_Reduce in the network. With our design, preliminary results show that we can achieve a 2.5× speedup for long messages and a 10× speedup for short messages over conventional clusters.