Partitioned vs Non Partitioned Graph Performance

[Samtruong]

Hello, I ran the Frog framework exp executable on the V100 and K40 GPU by changing -arch=sm_20 in the Makefile to support -arch=sm_70 and -arch=sm_35 respectively. I also enabled the non-partitioned option by uncommenting gpu_sssp_edge_loop in sssp.cu, gpu_bfs_edge_loop in bfs.cu and gpu_cc_edge_loop in cc.cu. The graph used in the experiment is com-youtube which can be found here: https://sparse.tamu.edu/SNAP/com-Youtube

This is the result of running ./exp

The left side of the graph is Frog running on V100, right side is Frog on K40. All of the runs are check PASS . My concern is: it looks like for all runs, the non-partitioned implementations outperform the partitioned implementation. I cannot produce results where the asynchronous model (with coloring as partitioning scheme) outperform the bulk-sync model. Please let me know if I am doing something wrong.

Thank you!