Decentralized Training of Foundation Models in Heterogeneous Environments

TL;DR

This paper explores decentralized training of large foundation models across heterogeneous, geo-distributed networks, proposing a scheduling algorithm that significantly improves training efficiency over traditional homogeneous data center setups.

Contribution

It introduces the first scheduling algorithm for large model training in decentralized, heterogeneous environments, optimizing task allocation across diverse GPU networks.

Findings

Achieves 4.8X faster training in geo-distributed settings compared to Megatron.

Develops a formal cost model and an evolutionary algorithm for optimal task scheduling.

Demonstrates effectiveness through extensive experiments on real-world network measurements.

Abstract

Training foundation models, such as GPT-3 and PaLM, can be extremely expensive, often involving tens of thousands of GPUs running continuously for months. These models are typically trained in specialized clusters featuring fast, homogeneous interconnects and using carefully designed software systems that support both data parallelism and model/pipeline parallelism. Such dedicated clusters can be costly and difficult to obtain. Can we instead leverage the much greater amount of decentralized, heterogeneous, and lower-bandwidth interconnected compute? Previous works examining the heterogeneous, decentralized setting focus on relatively small models that can be trained in a purely data parallel manner. State-of-the-art schemes for model parallel foundation model training, such as Megatron, only consider the homogeneous data center setting. In this paper, we present the first study of training large foundation models with model parallelism in a decentralized regime over a heterogeneous network. Our key technical contribution is a scheduling algorithm that allocates different computational "tasklets" in the training of foundation models to a group of decentralized GPU devices connected by a slow heterogeneous network. We provide a formal cost model and further propose an efficient evolutionary algorithm to find the optimal allocation strategy. We conduct extensive experiments that represent different scenarios for learning over geo-distributed devices simulated using real-world network measurements. In the most extreme case, across 8 different cities spanning 3 continents, our approach is 4.8X faster than prior state-of-the-art training systems (Megatron).