A Performance Study of UCX over InfiniBand
A Performance Study of UCX over InfiniBand
复制标题
InfiniBand 上 UCX 的性能研究
DOI：
10.1109/ccgrid.2017.149
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
P. Balaji
中科院分区：
文献类型：
--
作者：
Nikela Papadopoulou;Lena Oden;P. Balaji
UCX是一个开源通信框架，具有两级API设计，旨在满足大型超级计算系统的需求。较低级别的接口UCT给数据传输增加了最小的开销，但需要用户付出相当大的努力。高级接口UCP更易于使用，但会增加一些通信开销。这项工作的重点是图表的性能ofUCX在无限带宽，动机是使用UCX作为中间件的高级通信库。我们分析了两级设计引起的性能下降以及这些性能损失的来源。特别是，我们的目标UCP的基本功能，评估其性能在无限带宽，并分析与UCT和动词相比，开销的来源。我们提出并评估了一些修复措施，以尽量减少这些开销，以提高UCP的性能和可扩展性。
UCX is an open-source communication framework with a two-level APIdesign targeted at addressing the needs of large supercomputingsystems. The lower-level interface, UCT, adds minimal overhead todata transfer but requires considerable effort from the user. Thehigher-level interface, UCP, is easier to use, but adds some overheadto the communication. This work focuses on charting the performance ofUCX over InfiniBand, motivated by the usage of UCX as middleware forhigh-level communication libraries. We analyze performanceshortcomings that stem from the two-level design and the sources ofthese performance losses. In particular, we target basic functions ofUCP, evaluate their performance over InfiniBand, and analyze sourcesof overheads compared with UCT and Verbs. We propose and evaluatesome fixes to minimize these overheads, in order to enhance UCPperformance and scalability.