A Software-Hardware Co-Optimized Toolkit for Deep Reinforcement Learning on Heterogeneous Platforms

TL;DR

This paper presents PEARL, a versatile toolkit that optimizes deep reinforcement learning on heterogeneous hardware platforms, improving performance and power efficiency through hardware-agnostic protocols and automatic device assignment.

Contribution

PEARL introduces a hardware-agnostic training protocol, DRL-specific scheduling, and automatic task-device optimization for heterogeneous platforms, advancing the efficiency of DRL implementations.

Findings

Outperforms state-of-the-art CPU-GPU libraries in throughput by up to 2.1×.

Achieves power efficiency improvements of up to 3.4×.

Demonstrates effectiveness on DQN and DDPG algorithms across diverse platforms.

Abstract

Deep Reinforcement Learning (DRL) is vital in various AI applications. DRL algorithms comprise diverse compute kernels, which may not be simultaneously optimized using a homogeneous architecture. However, even with available heterogeneous architectures, optimizing DRL performance remains a challenge due to the complexity of hardware and programming models employed in modern data centers. To address this, we introduce PEARL, a toolkit for composing parallel DRL systems on heterogeneous platforms consisting of general-purpose processors (CPUs) and accelerators (GPUs, FPGAs). Our innovations include: 1. A general training protocol agnostic of the underlying hardware, enabling portable implementations across various processors and accelerators. 2. Incorporation of DRL-specific scheduling optimizations within the protocol, facilitating parallelized training and enhancing the overall system performance. 3. High-level API for productive development using the toolkit. 4. Automatic optimization of DRL task-to-device assignments through performance estimation, supporting various optimization metrics including throughput and power efficiency. We showcase our toolkit through experimentation with two widely used DRL algorithms, DQN and DDPG, on two diverse heterogeneous platforms. The generated implementations outperform state-of-the-art libraries for CPU-GPU platforms by throughput improvements of up to 2.1$\times$ and power efficiency improvements of up to 3.4$\times$.