Hi, I have some iterative program written in Spark and have been tested under a snapshot version of spark 0.8 before. After I ported it to the 0.8 release version, I see performance drops in large datasets. I am wondering if there is any clue?
I monitored the number of partitions on each machine (by looking at DAGScheduler.getCacheLocs). I observed that some machine may have 30 partitions in the previous iteration while only have < 10 partitions in the next iterations. This is something I didn't observed in the older version. Thus I am wondering if the release version would do task stealing more aggressively (for a better dynamic load balance?) Thank you! Best Regards, Wenlei