Introduction to Distributed DL Training

Modern Deep Learning systems heavily rely on distributed training over customized high-performance accelerator (e.g., TPU, GPU)-based hardware platforms connected via high-performance interconnects (e.g., NVLink). Examples today include NVIDIA’s DGX-2, Google’s Cloud TPU and Facebook’s Zion. Deep Neural Network (DNN) training involves a complex interplay between the DNN model architecture, parallelization strategy, scheduling strategy, collective communication algorithm, network topology, and the accelerator endpoint. Collective communications (e.g., All-Reduce, All-to-All, Reduce-Scatter, All-Gather) are initiated at different phases for different parallelism approaches — and play a crucial role in overall runtime, if not hidden efficiently behind compute. This problem becomes paramount as recent models for NLP such as GPT-3 and Recommendations such as DLRM have billions to trillions of parameters and need to be scaled across tens to hundreds to thousands of accelerator nodes. As innovation in AI/ML models continues to grow at an accelerated rate, there is a need for a comprehensive methodology to understand and navigate this complex design-space to (i) architect future platforms and (ii) develop novel parallelism schemes to support efficient training of future DNN models.

As an ongoing collaboration between Intel, Facebook and Georgia Tech, we have been jointly developing a detailed cycle-accurate distributed training simulator called ASTRA-sim. ASTRA-sim models the co-design space described above and schedules the compute-communication interactions from distributed training over plug-and-play compute and network simulators. It enables a systematic study of bottlenecks at the software and hardware level for scaling training. It also enables end-to-end design-space exploration for running large DNN models over future training platforms. Currently, ASTRA-sim supports two compute models (roofline and SCALE-sim, a Google TPU-like simulator) and several network models (analytical network, Garnet from gem5, and NS3) to go from simple analytical to detailed cycle-accurate simulation of large-scale training platforms. In this tutorial, we will educate the research community about the challenges in the emerging domain of distributed training, demonstrate the capabilities of ASTRA-sim with examples and discuss ongoing development efforts.