Multi-GPU Dynamic Scheduling Pitfalls occupancy: ensure make the minimum number of blocks to satisfy the utilization requirement 2048/128 = 16 at leat, but still tail-effects... unified-memory-prefetch: ensure much more regular page access pattern, do not split into too much tasks (maybe just 8) File Organization file description ../scan_xp_common.cu scan-xp common algorithmic functions, directly copy and paste into separate files ../cuda_utils/set_inter_device_functions.cuh set-intersection scalar device functions ../cuda_utils/cuda_util.cuh cuda utilities file description scan_xp.cu warning: depracated: single-gpu implementation scan_xp_multipass.cu single-gpu multiple-pass for snap-friendster and larger datasets scan_xp_multigpu.cu warning: depracated: a special case of scan_xp_multigpu_multipass.cu scan_xp_multigpu_multipass.cu multi-gpu multiple-pass implementation for snap-friendster and larger datasets Compile (CUDA-ALL, with MergeBased and Baseline) see ../CMakeLists.txt attention: first time make -j may report error, then make -j again to build all cuda executables cmake .. -DEXTRA_INCLUDE_DIR=/homes/ywangby/workspace/yche/local/include make -j make -j Run CUDA_VISIBLE_DEVICES=1 ./scan-xp-cuda-bitmap-warp-per-vertex-multi-gpu-multi-pass /export/data/set-inter-datasets/data/dataset/webgraph_twitter/rev_deg 0.2 5 56