Workshop: The 9th Annual Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM)
Authors: Dalton Beauchesne (Unviersity of Saskatchewan) and Fahim Hamid, Kyle Klenk, and Raymond Spiteri (University of Saskatchewan)
Abstract: The C++ Actor Framework (CAF) once provided OpenCL-based GPU acceleration within the actor model, but this capability is no longer supported. This paper presents CUDA-Enabled Actors, a CUDA-based extension that re-establishes a practical path to GPU execution in CAF while preserving an actor-oriented programming interface. CUDA support is implemented through a thread-safe command runner that serves as the execution interface for launching CUDA kernels through an actor-oriented API and gives users finer control over host--device memory transfers. The resulting design hides much of the complexity of GPU programming while retaining the modularity and abstraction of the actor model. Experiments demonstrate that CUDA-Enabled Actors provides actor-oriented GPU execution with modest overhead relative to native CUDA, supports execution across multiple GPUs, enables recovery from worker failures, and improves resource use for irregular workloads, establishing it as a practical foundation for GPU-enabled applications in CAF.
Back to The 9th Annual Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM) Archive Listing
Back to Full Workshop Archive Listing