Improving All-to-Many Personalized Communication in Two-Phase I/O
SC, 2020
This paper generalizes the spread-out algorithm and adopts communication throttling to reduce the straggler and contention effects of the all-to-many personalized communication used in two-phase MPI-IO request redistribution, and shows on Cori and Summit that tailoring the communication kernel to different I/O patterns can improve end-to-end performance by up to 10x over default MPI-IO.
BibTeX
@inproceedings{kang2020improving,
title={Improving all-to-many personalized communication in two-phase i/o},
author={Kang, Qiao and Ross, Robert and Latham, Robert and Lee, Sunwoo and Agrawal, Ankit and Choudhary, Alok and Liao, Wei-keng},
booktitle={SC20: International Conference for High Performance Computing, Networking, Storage and Analysis},
pages={1--13},
year={2020},
organization={IEEE}
}