Abstract-Most recent HPC platforms have heterogeneous nodes composed of multi-core CPUs and accelerators, like GPUs. Programming such nodes is typically based on a combination of OpenMP and CUDA/OpenCL codes; scheduling relies on a static partitioning and cost model.We present the XKaapi runtime system for data-flow task programming on multi-CPU and multi-GPU architectures, which supports a data-flow task model and a localityaware work stealing scheduler. XKaapi enables task multiimplementation on CPU or GPU and multi-level parallelism with different grain sizes. We show performance results on two dense linear algebra kernels, matrix product (GEMM) and Cholesky factorization (POTRF), to evaluate XKaapi on a heterogeneous architecture composed of two hexa-core CPUs and eight NVIDIA Fermi GPUs.Our conclusion is two-fold. First, fine grained parallelism and online scheduling achieve performance results as good as static strategies, and in most cases outperform them. This is due to an improved work stealing strategy that includes locality information; a very light implementation of the tasks in XKaapi; and an optimized search for ready tasks. Next, the multi-level parallelism on multiple CPUs and GPUs enabled by XKaapi led to a highly efficient Cholesky factorization. Using eight NVIDIA Fermi GPUs and four CPUs, we measure up to 2.43 TFlop/s on double precision matrix product and 1.79 TFlop/s on Cholesky factorization; and respectively 5.09 TFlop/s and 3.92 TFlop/s in single precision.

International audienceToday, it is possible to associate multiple CPUs and multiple GPUs in a single shared memory architecture. Using these resources efficiently in a seamless way is a challenging issue. In this paper, we propose a parallelization scheme for dynamically balancing work load between multiple CPUs and GPUs. Most tasks have a CPU and GPU implementation, so they can be executed on any processing unit. We rely on a two level scheduling associating a traditional task graph partitioning and a work stealing guided by processor affinity and heterogeneity. These criteria are intended to limit inefficient task migrations between GPUs, the cost of memory transfers being high, and to favor mapping small tasks on CPUs and large ones on GPUs to take advantage of heterogeneity. This scheme has been implemented to support the SOFA physics simulation engine. Experiments show that we can reach speedups of 22 with 4 GPUs and 29 with 4 CPU cores and 4 GPUs. CPUs unload GPUs from small tasks making these GPUs more efficient, leading to a "cooperative speedup" greater than the sum of the speedups separatly obtained on 4 GPUs and 4 CPUs

This paper introduces FlowVR, a middleware dedicated to virtual reality applications distributed on clusters or grid environments. FlowVR supports coupling of heterogeneous parallel codes and is component oriented to favor code reuse. While classical communication paradigms focus on either a synchronous approach (FIFO channels) or an asynchronous one (sampling), FlowVR enables a large range of intermediate policies to better balance the application performance between levels of details, latencies and refresh rates.