Github user vrilleup commented on the pull request: https://github.com/apache/spark/pull/964#issuecomment-48247731 @yangliuyu sorry for the delayed reply. I think there has to be enough parallelism to fully power Spark. If you have 16 cores, why not get more executors or assign more cores per executor to allow more tasks running at the same time? I am not sure about the scheduler delay and GC time. In my test case with n = 300k, scheduler delay is 0.3s (max 1.0s), and GC takes no time (occasionally takes <100ms in some tasks). Do you have enough memory to hold all the singular vectors? n * (6 * k + 4) doubles need to fit in the master node memory for ARPACK, and additional memory is required to cache the RDD.
--- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastruct...@apache.org or file a JIRA ticket with INFRA. ---