Improving data transfer for model coupling

changshui zhang,litao liu,geng yang,r li,bin wang

DOI: https://doi.org/10.5194/gmdd-8-8981-2015

2015-01-01

Geoscientific Model Development Discussions

Abstract:12 Data transfer means transferring data fields between two component models or rearranging 13 data fields among processes of the same component model. It is a fundamental and most 14 frequently used operation of a coupler. Most versions of state-of-the-art couplers currently use 15 an implementation based on the point-to-point (P2P) communication of the Message Passing 16 Interface (MPI) (refer such an implementation as “P2P implementation” for short). In this 17 paper, we reveal the drawbacks of the P2P implementation, including low communication 18 bandwidth due to small message size, variable and big number of MPI messages, as well as 19 network contention. To overcome these drawbacks, we propose a butterfly implementation 20 for data transfer. Although the butterfly implementation can outperform the P2P 21 implementation in many cases, it degrades the performance in some cases because the total 22 message size transferred by the butterfly implementation is larger than the total message size 23 transferred by the P2P implementation. To further improve data transfer, we design and 24 implement an adaptive data transfer library that combines the advantages of both butterfly 25 implementation and P2P implementation. Performance evaluation shows that the adaptive 26 data transfer library significantly improves the performance of data transfer in most cases, and 27 does not decrease the performance in any cases. Now, the adaptive data transfer library is 28