Towards Universal Performance Modeling for Machine Learning Training on Multi-GPU Platforms

TL;DR

This paper presents a comprehensive performance modeling framework for multi-GPU machine learning training, accurately predicting training times and aiding configuration choices across diverse workloads and hardware setups.

Contribution

It introduces a data-distribution-aware performance model and communication prediction techniques, extending prior single-GPU work to multi-GPU environments with high accuracy.

Findings

Predicts per-iteration training time with 5.21% error for DLRM models.

Generalizes well to Transformer-based NLP models with 3.00% error.

Enables quick selection of optimal embedding sharding configurations with 85% success rate.

Abstract

Characterizing and predicting the training performance of modern machine learning (ML) workloads on compute systems with compute and communication spread between CPUs, GPUs, and network devices is not only the key to optimization and planning but also a complex goal to achieve. The primary challenges include the complexity of synchronization and load balancing between CPUs and GPUs, the variance in input data distribution, and the use of different communication devices and topologies (e.g., NVLink, PCIe, network cards) that connect multiple compute devices, coupled with the desire for flexible training configurations. Built on top of our prior work for single-GPU platforms, we address these challenges and enable multi-GPU performance modeling by incorporating (1) data-distribution-aware performance models for embedding table lookup, and (2) data movement prediction of communication collectives, into our upgraded performance modeling pipeline equipped with inter-and intra-rank synchronization for ML workloads trained on multi-GPU platforms. Beyond accurately predicting the per-iteration training time of DLRM models with random configurations with a geomean error of 5.21% on two multi-GPU platforms, our prediction pipeline generalizes well to other types of ML workloads, such as Transformer-based NLP models with a geomean error of 3.00%. Moreover, even without actually running ML workloads like DLRMs on the hardware, it is capable of generating insights such as quickly selecting the fastest embedding table sharding configuration (with a success rate of 85%).