dOpenCL: Towards Uniform Programming for Distributed Systems with Multi-Cores and GPUs

Modern computer systems are becoming distributed and heterogeneous by comprising multi-core CPUs, GPUs, and other accelerators. However, to program such systems, the user currently has to use a combination of several programming models (e.g., MPI with OpenCL or CUDA), which is difficult and error-prone. We present dOpenCL (distributed OpenCL) – a uniform approach to programming distributed systems with accelerators. Our approach is based on the OpenCL standard and it allows the user to run existing OpenCL applications unmodified in a heterogeneous distributed environment. The dOpenCL system also supports transparent execution of multiple OpenCL applications in one distributed, multi-user environment. We describe dOpenCL as an implementation of the OpenCL programming model on distributed systems, and we experimentally compare the performance of dOpenCL with MPI+OpenCL and standard OpenCL implementations.