taking advantage of the spare compute to reduce the memory bandwidth usage.

Do we know why see a slight improvement for PageRank with compression on a one core system? Is it just because we are doing less memory transactions?

This result indicates that it is important to understand the system that the program executes on when trying to decide which approach to use

Explanation of the two graphs: The graph processing is unlikely to be bandwidth bound when running on only one core. However, when scaling up to 40 cores, as the memory system is the same, it is likely to be bandwidth bound. So that graph compression, which will burn computation to reduce memory BW, can run faster on 40 cores.

Performance boost from compression is more pronounced on systems with even more severe problem of bandwidth bound, i.e. lower Arithmetic Intensity. With multi-core and multiple times more compute power, bandwidth bottleneck is even more problematic