A wide range of reinforcement learning (RL) algorithms have been proposed, in which agents learn from interactions with a simulated environment. Executing such RL training loops is computationally expensive, but current RL systems fail to support the training loops of different RL algorithms efficiently on GPU clusters: they either hard-code algorithm-specific strategies for parallelization and distribution; or they accelerate only parts of the computation on GPUs (e.g., DNN policy updates). We observe that current systems lack an abstraction that decouples the definition of an RL algorithm from its strategy for distributed execution.
We describe MSRL, a distributed RL training system
that uses the new abstraction of a fragmented dataflow
graph (FDG) to execute RL algorithms in a flexible way.
An FDG is a heterogeneous dataflow representation of an
RL algorithm, which maps functions from the RL training
loop to independent parallel dataflow fragments. Fragments
account for the diverse nature of RL algorithms: each fragment can execute on a different device using its own low-level dataflow implementation, e.g., an operator graph of a DNN engine, a CUDA GPU kernel, or a multi-threaded CPU process. At deployment time, a distribution policy governs how fragments are mapped to devices, without changes to the algorithm implementation. Our experiments show that MSRL exposes trade-offs between different execution strategies, while surpassing the performance of existing RL systems.