RUNTIME MESSAGE COMPRESSION TO IMPROVE THE COUPLING IN EARTH SYSTEM MODELING

摘要

As is extensively used in today's Earth System Modeling community, the coupler is a critical component aimed to link two or more large-scale numerical simulations called models. One basic function of the coupler is the parallel data transfer, which is used to exchange coupled data between models. With the advent of next-generation models, higher horizontal resolutions and more new coupling requirements will cause a large amount of communication to the coupler. Using runtime compression can effectively enhance the communication performance by reducing the data volume for transfer. In this paper, we study how to improve coupling performance by using the compression methods. Through two sets of tests using real coupled data, we characterize the performance of data transfer using compression on the clusters interconnected by Gigabit Ethernet and Infiniband, respectively. The results show that using LZO and two specific compressions proposed by us can achieve at least 1.05 to 5.1 fold speedups on the cluster with Gigabit network. The results also show that when the communication contention exists, using compression can achieve 1.3 to 3.5 fold speedups for the coupling between some models in an optimal computing environment.