Efficient Communication Acceleration for Next-GenScale-up Deep Learning Training Platforms
share
Deep Learning (DL) training platforms are built by interconnecting multiple DL accelerators (e.g., GPU/TPU) via fast, customized interconnects. As the size of DL models and the compute efficiency of the accelerators has continued to increase, there has also been a corresponding steady increase in the bandwidth of these interconnects.Systems today provide 100s of gigabytes (GBs) of inter-connect bandwidth via a mix of solutions such as Multi-Chip packaging modules (MCM) and proprietary interconnects(e.g., NVlink) that together from the scale-up network of accelerators. However, as we identify in this work, a significant portion of this bandwidth goes under-utilized. This is because(i) using compute cores for executing collective operations such as all-reduce decreases overall compute efficiency, and(ii) there is memory bandwidth contention between the accesses for arithmetic operations vs those for collectives, and(iii) there are significant internal bus congestions that increase the latency of communication operations. To address this challenge, we propose a novel microarchitecture, calledAccelerator Collectives Engine(ACE), forDL collective communication offload. ACE is a smart net-work interface (NIC) tuned to cope with the high-bandwidth and low latency requirements of scale-up networks and is able to efficiently drive the various scale-up network systems(e.g. switch-based or point-to-point topologies). We evaluate the benefits of the ACE with micro-benchmarks (e.g. single collective performance) and popular DL models using an end-to-end DL training simulator. For modern DL workloads, ACE on average increases the net-work bandwidth utilization by 1.97X, resulting in 2.71X and 1.44X speedup in iteration time for ResNet-50 and GNMT, respectively.
READ FULL TEXT