Dynamic performance prediction for chunk-wise parallelization on heterogeneous CPU/GPU systems

摘要

Many aspects of heterogeneity in multicores such as performance variation may affect the overall execution time and cores efficiency. An effective mapping should support this variation. A complex challenge is cores load balancing to minimize the program makespan. In this context, we propose a predictive approach using iterations chunking at runtime allowing parallel code adaptation to heterogeneous systems containing GPU and heterogeneous CPUs. Our approach is based on thread pinning and performance detection at runtime. From a parallel program, we run a chunk on the GPU device. Another chunk is running on CPU using a first mapping assuming homogeneous cores. Then, performance assessment would correct mapping by speculating the future GPU and core's state. The new mapping would be then applied to a new chunk for further evaluation and prediction. This process would stop when the program is fully executed or when judging that chunking is no longer effective.