Инструменты анализа и разработки эффективного кода для параллельных архитектур

Труды Института системного программирования РАН. 2014. Т. 26. № 1. С. 357–374.

Монаков А. В., Платонов В. А., Avetisyan A.

The article proposes methods for supporting development of efficient programs for modern parallel architectures, including hybrid systems. First, specialized profiling methods designed for programmers tasked with parallelizing existing code are proposed. The first method is loop-based profiling via source-level instrumentation done with Coccinelle tool. The second method is memory reuse distance estimation via virtual memory protection mechanism and manual instrumentation. The third method is cache miss and false sharing estimation by collecting a partial trace of memory accesses using compiler instrumentation and estimating cache behavior in postprocessing based on the trace and a cache model. Second, the problem of automatic parallel code generation for hybrid architectures is discussed. Our approach is to generate OpenCL code from parallel loop nests based on GRAPHITE infrastructure in the GCC compiler. Finally, in cases where achieving high efficiency on hybrid systems requires significant rework of data structures or algorithms, one can employ auto-tuning to specialize for specific input data and hardware at run time. This is demonstrated on the problem of optimizing sparse matrix-vector multiplication for GPUs and its use for accelerating linear system solving in OpenFOAM CFD package. We propose a variant of “sliced ELLPACK” sparse matrix storage format with special treatment for small horizontal or diagonal blocks, where the exact parameters of matrix structure and GPU kernel launch should be automatically tuned at runtime for the specific matrix and GPU hardware.

Keywords: CUDA, исследование и оптимизация программ, OpenCL, OpenFOAM, профилирование, разреженные матрицы