high performance computing on graphics processing units: hgpu.org

High performance architectures are increasingly heterogeneous with shared and distributed memory components, and accelerators like GPUs. Programming such architectures is complicated and performance portability is a major issue as the architectures evolve. This thesis explores the potential for algorithmic skeletons integrating a dynamically parametrised static cost model, to deliver portable performance for mostly regular data parallel programs on heterogeneous archi- tectures. The rst contribution of this thesis is to address the challenges of program- ming heterogeneous architectures by providing two skeleton-based programming libraries: i.e. HWSkel for heterogeneous multicore clusters and GPU-HWSkel that enables GPUs to be exploited as general purpose multi-processor devices. Both libraries provide heterogeneous data parallel algorithmic skeletons including hMap, hMapAll, hReduce, hMapReduce, and hMapReduceAll. The second contribution is the development of cost models for workload dis- tribution. First, we construct an architectural cost model (CM1) to optimise overall processing time for HWSkel heterogeneous skeletons on a heterogeneous system composed of networks of arbitrary numbers of nodes, each with an ar- bitrary number of cores sharing arbitrary amounts of memory. The cost model characterises the components of the architecture by the number of cores, clock speed, and crucially the size of the L2 cache. Second, we extend the HWSkel cost model (CM1) to account for GPU performance. The extended cost model (CM2) is used in the GPU-HWSkel library to automatically nd a good distribution for both a single heterogeneous multicore/GPU node, and clusters of heteroge- neous multicore/GPU nodes. Experiments are carried out on three heterogeneous multicore clusters, four heterogeneous multicore/GPU clusters, and three single heterogeneous multicore/GPU nodes. The results of experimental evaluations for four data parallel benchmarks, i.e. sumEuler, Image matching, Fibonacci, and Matrix Multiplication, show that our combined heterogeneous skeletons and cost models can make good use of resources in heterogeneous systems. Moreover using cores together with a GPU in the same host can deliver good performance either on a single node or on multiple node architectures.