Introduction || What is DDP || Single-Node Multi-GPU Training || Fault Tolerance || Multi-Node training || minGPT Training

What is Distributed Data Parallel (DDP)

Authors: Suraj Subramanian

What you will learn

• How DDP works under the hood

• What is the DistributedSampler

• How gradients are synchronized across GPUs

Prerequisites

• Familiarity with basic non-distributed training in PyTorch

Follow along with the video below or on youtube.

This tutorial is a gentle introduction to PyTorch DistributedDataParallel (DDP) which enables data parallel training in PyTorch. Data parallelism is a way to process multiple data batches across multiple devices simultaneously to achieve better performance. In PyTorch, the DistributedSampler ensures each device gets a non-overlapping input batch. The model is replicated on all the devices; each replica calculates gradients and simultaneously synchronizes with the others using the ring all-reduce algorithm.

Why you should prefer DDP over DataParallel (DP)

DataParallel is an older approach to data parallelism. DP is trivially simple (with just one extra line of code) but it is much less performant. DDP improves upon the architecture in a few ways:

DataParallel	DistributedDataParallel
More overhead; model is replicated and destroyed at each forward pass	Model is replicated only once
Only supports single-node parallelism	Supports scaling to multiple machines
Slower; uses multithreading on a single process and runs into Global Interpreter Lock (GIL) contention	Faster (no GIL contention) because it uses multiprocessing

Further Reading

• Multi-GPU training with DDP (next tutorial in this series)

• DDP API

• DDP Internal Design