Multi-level Temporal Blocking for Stencil Computation for Memory Hierarchy on TSUBAME2.5

概要

論文の詳細を見る
The domain of the stencil computation is limited by the memory capacity of GPUs on a GPU cluster. As the domain grows to cope with higher accuracy requirements, more GPUs need to be employed to extend the memory capacity. In this paper, we propose new methods which apply temporal blocking method to device memory and registers of a set of GPUs to allow computations on the domain that is bigger than the memory capacity of GPUs while maintaining high performance on TSUBAME2.5. We also analyze the parameters and performance differences between TSUBAME2.0 and TSUBAME2.5 to apply our methods to wide range GPU clusters.
2014-02-24