User 90 | 3/24/2014, 7:47:12 AM
How does graph diameter and degree-skewness affect the performance of applications on GraphLab? Are there any specific optimizations that are used for high/low diameter graphs and maybe power-law graphs?
I am looking at following data for 2 programs in graph-analytics toolkit. LiveJournal is 4.8M vertices & 68.9M edges whereas RoadNet-CA is 1.9M vertices & 5.5M edges.
PageRank: LiveJournal: 295 sec (sync) ;; 716 sec (async) RoadNet-CA: 129 sec (sync) ;; 604sec (async)
SSSP: LiveJournal: 15 sec (sync) ;; 730 sec (async) RoadNet-CA: 82 sec (sync) ;; 60 sec (async)
Why does RoadNet-CA perform reverse in SSSP (async faster than sync) compared to PageRank (sync way faster than async)?