Lazy Allocation and Transfer Fusion Optimization for GPU-based Heterogeneous Systems

We present two memory optimization techniques which improve the efficiency of data transfer over PCIe bus for GPU-based heterogeneous systems, namely lazy allocation and transfer fusion optimization. Both are based on merging data transfers so that less overhead is incurred, thereby increasing transfer throughput and making accelerator usage profitable also for smaller operand sizes. We provide the design and prototype implementation of the two techniques in CUDA. Microbench-marking results show that especially for smaller and medium-sized operands significant speedups can be achieved. We also prove that our transfer fusion optimization algorithm is optimal.