Open hucheng opened 9 years ago
BBR has been tested. Experimental result suggests that BBR is slower than DBH when both partitioners work, but when the topic number is large or it has too many partitions, DBH can't work while BBR works well. Label propagation partitioner has basically close performance with BBR, but it has two weaknesses: 1. It is an iterative algorithm so it need re-partition again and again so it is slow; 2. It needs transfer a P * P matrix in each iteration which P is the partition number. When P is large, the rdd.collect() call can be very slow and cause the driver OOM.
Good.
Can you make these part of PartitionStrategy
instead of implementing in a separate package? Seems like they belong in the graphx2 package...
There are four partitioning strategies in GraphX:
Besides, we also implemented: