Self-adaptive and Topology-aware Broadcast Leveraging Collective Offload on Tianhe Express Interconnect

Chongshan Liang

doi:10.1016/j.jpdc.2026.105274

Self-adaptive and Topology-aware Broadcast Leveraging Collective Offload on Tianhe Express Interconnect

2026-04-14

0

PRE

AI

C

Chongshan Liang *

Y

Yi Dai

X

Xingyun Qi

J

Jiaqing Xu

M

Min Xie

D

Dongsheng Li

W

Weixia Xu

DOI：10.1016/j.jpdc.2026.105274

Original

Original request for help

Save

Abstract

En 中文

• It presents a hardware-software co-design that completely offloads both blocking and non-blocking broadcast operations to the NIC by taking full advantage of the NIC’s collective offload features. • It develops a topology-aware optimization for collective communication. • It proposes a hardware-based non-blocking MPI_Ibcast mechanism. • Extensive experiments conducted on production HPC supercomputers show that the proposed approach can significantly speed up the broadcast at both microbenchmark and application level.

Keywords:

NIC

collective offload

non-blocking broadcast

topology-aware optimization

MPI_Ibcast

Self-adaptive and Topology-aware Broadcast Leveraging Collective Offload on Tianhe Express Interconnect

Abstract

Journal

Organization

Cited Papers

Citing Papers