Efficient and Eventually Consistent Collective Operations

TL;DR

This paper introduces an efficient, eventually consistent approach to collective operations in parallel computing, reducing communication overhead and improving performance for ML/DL and HPC applications, especially in strong scaling scenarios.

Contribution

It proposes a novel design for eventually consistent collectives, optimizing Broadcast and Reduce, and integrates classic collectives into GASPI, demonstrating promising preliminary performance gains.

Findings

Significant improvements in Allreduce and AlltoAll performance

Reduced communication in Broadcast and Reduce operations

Enhanced GASPI ecosystem with new collective implementations

Abstract

Collective operations are common features of parallel programming models that are frequently used in High-Performance (HPC) and machine/ deep learning (ML/ DL) applications. In strong scaling scenarios, collective operations can negatively impact the overall application performance: with the increase in core count, the load per rank decreases, while the time spent in collective operations increases logarithmically. In this article, we propose a design for eventually consistent collectives suitable for ML/ DL computations by reducing communication in Broadcast and Reduce, as well as by exploring the Stale Synchronous Parallel (SSP) synchronization model for the Allreduce collective. Moreover, we also enrich the GASPI ecosystem with frequently used classic/ consistent collective operations -- such as Allreduce for large messages and AlltoAll used in an HPC code. Our implementations show promising preliminary results with significant improvements, especially for Allreduce and AlltoAll, compared to the vendor-provided MPI alternatives.