Programming models and techniques to exploit parallelism in accelerators, such as GPUs, are different from those used in traditional parallel models for shared- or distributed-memory systems. It is a challenge to blend different programming models to coordinate and exploit devices with very different characteristics and computation powers. This paper presents a new extensible framework model to encapsulate run-time decisions related to data partition, granularity, load balance, synchronization, and communication for systems including assorted GPUs. Thus, the main parallel code becomes independent of them, using internal topology and system information to transparently adapt the computation to the system. The programmer can develop specific functions for each architecture, or use existent specialized library functions for different CPU-core or GPU architectures. The high-level coordination is expressed using a programming model built on top of message-passing, providing portability across distributed- or shared-memory systems. We show with an example how to produce a parallel code that can be used to efficiently run on systems ranging from a Beowulf cluster to a machine with mixed GPUs. Our experimental results show how the run-time system, guided by hints about the computational-power ratios of different devices, can automatically part and distribute large computations across heterogeneous systems, improving the overall performance.
