FreeScale: Distributed Training for Sequence Recommendation Models with Minimal Scaling Cost

TL;DR

FreeScale is a distributed training method for sequence recommendation models that significantly reduces computational inefficiencies and resource under-utilization on large GPU clusters.

Contribution

It introduces load balancing, communication overlapping, and SM-Free techniques to address stragglers and communication bottlenecks in large-scale training.

Findings

Achieves up to 90.3% reduction in computational bubbles.

Effectively mitigates stragglers and communication delays.

Improves GPU resource utilization in real-world workloads.

Abstract

Modern industrial Deep Learning Recommendation Models typically extract user preferences through the analysis of sequential interaction histories, subsequently generating predictions based on these derived interests. The inherent heterogeneity in data characteristics frequently result in substantial under-utilization of computational resources during large-scale training, primarily due to computational bubbles caused by severe stragglers and slow blocking communications. This paper introduces FreeScale, a solution designed to (1) mitigate the straggler problem through meticulously load balanced input samples (2) minimize the blocking communication by overlapping prioritized embedding communications with computations (3) resolve the GPU resource competition during computation and communication overlapping by communicating through SM-Free techniques. Empirical evaluation demonstrates that FreeScale achieves up to 90.3% reduction in computational bubbles when applied to real-world workloads running on 256 H100 GPUs.