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 audienceThe race for Exascale computing has naturally led the current technologies to converge to multi-CPU/multi-GPU computers, based on thousands of CPUs and GPUs interconnected by PCI-Express buses or interconnection networks. To exploit this high computing power, programmers have to solve the issue of scheduling parallel programs on hybrid architectures. And, since the performance of a GPU increases at a much faster rate than the throughput of a PCI bus, data transfers must be managed efficiently by the scheduler. This paper targets multi-GPU compute nodes, where several GPUs are connected to the same machine. To overcome the data transfer limitations on such platforms, the available softwares compute, usually before the execution, a mapping of the tasks that respects their dependencies and minimizes the global data transfers. Such an approach is too rigid and it cannot adapt the execution to possible variations of the system or to the application's load. We propose a solution that is orthogonal to the above mentioned: extensions of the XKaapi software stack that enable to exploit full performance of a multi-GPUs system through asynchronous GPU tasks. XKaapi schedules tasks by using a standard Work Stealing algorithm and the runtime efficiently exploits concurrent GPU operations. The runtime extensions make it possible to overlap the data transfers and the task executions on current generation of GPUs. We demonstrate that the overlapping capability is at least as important as computing a scheduling decision to reduce completion time of a parallel program. Our experiments on two dense linear algebra problems (Matrix Product and Cholesky factorization) show that our solution is highly competitive with other softwares based on static scheduling. Moreover, we are able to sustain the peak performance (310 GFlop/s) on DGEMM, even for matrices that cannot be stored entirely in one GPU memory. With eight GPUs, we archive a speed-up of 6.74 with respect to single-GPU. The performance of our Cholesky factorization, with more complex dependencies between tasks, outperforms the state of the art single-GPU MAGMA code

Abstract. Efficient implementations of parallel applications on heterogeneous hybrid architectures require a careful balance between computations and communications with accelerator devices. Even if most of the communication time can be overlapped by computations, it is essential to reduce the total volume of communicated data. The literature therefore abounds with ad hoc methods to reach that balance, but that are architecture and application dependent. We propose here a generic mechanism to automatically optimize the scheduling between CPUs and GPUs, and compare two strategies within this mechanism: the classical Heterogeneous Earliest Finish Time (HEFT) algorithm and our new, parametrized, Distributed Affinity Dual Approximation algorithm (DADA), which consists in grouping the tasks by affinity before running a fast dual approximation. We ran experiments on a heterogeneous parallel machine with six CPU cores and eight NVIDIA Fermi GPUs. Three standard dense linear algebra kernels from the PLASMA library have been ported on top of the XKaapi runtime. We report their performances. It results that HEFT and DADA perform well for various experimental conditions, but that DADA performs better for larger systems and number of GPUs, and, in most cases, generates much lower data transfers than HEFT to achieve the same performance.

International audienceIn this paper, we present a comparison of scheduling strategies for heterogeneous multi-CPU and multi-GPU architectures. We designed and evaluated four scheduling strategies on top of XKaapi runtime: work stealing, data-aware work stealing, locality-aware work stealing, and Heterogeneous Earliest-Finish-Time (HEFT). On a heterogeneous architecture with 12 CPUs and 8 GPUs, we analysed our scheduling strategies with four benchmarks: a BLAS-1 AXPY vector operation, a Jacobi 2D iterative computation, and two linear algebra algorithms Cholesky and LU. We conclude that the use of work stealing may be efficient if task annotations are given along with a data locality strategy. Furthermore, our experimental results suggests that HEFT scheduling performs better on applications with very regular computations and low data locality

In the last ten years, GPUs have dominated the market considering the computing/power metric and numerous research works have provided Basic Linear Algebra Subprograms implementations accelerated on GPUs. Several software libraries have been developed for exploiting performance of systems with accelerators, but the real performance may be far from the platform peak performance. This paper presents XKBlas that aims to improve performance of BLAS-3 kernels on multi-GPU systems. At low level, we model computation as a set of tasks accessing data on different resources. At high level, the API design favors non-blocking calls as uniform concept to overlap latency, even by fine grain computation. Unit benchmark of BLAS-3 kernels showed that XKBlas outperformed most implementations including the overhead of dynamic task's creation and scheduling. XKBlas outperformed BLAS implementations such as cuBLAS-XT, PaRSEC, BLASX and Chameleon/StarPU.