[ https://issues.apache.org/jira/browse/GIRAPH-306?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13437607#comment-13437607 ]
Avery Ching commented on GIRAPH-306: ------------------------------------ >Yeah that was the impression I had too. Just to clarify, as of the recent >Netty upgrades + this one, we are in no way >attempting to handle worker >restarts with any grace right? This is all purely connection reliability for >healthy worker nodes? Yeah, this is purely for reliability of connections and requests, nothing else. >I am having a lot more trouble scaling out to more workers than I used to. I >know you guys had mentioned this, but I have >not been testing again until the >last few days and its definitely gotten trickier, not the least of which >because I'm >having trouble getting logs to see what happened during a fail. I >don't have dumps I saved from those jobs, but if I see >more I will put them >here. Here's a trick you can try. Add -Dmapred.map.max.attempts=1 to ensure that any failure will fail the job. Then you can see the logs for the failed task and try to figure out what the problem is. >Mostly the logs I get are reconnection logs after reincarnation in which they >all fail (of course) and no logs for the >failed portion of the run that >triggered the worker to reincarnate. The above should help us narrow down your problem. =) > Netty requests should be reliable and implement exactly once semantics > ---------------------------------------------------------------------- > > Key: GIRAPH-306 > URL: https://issues.apache.org/jira/browse/GIRAPH-306 > Project: Giraph > Issue Type: Improvement > Reporter: Avery Ching > Priority: Critical > Attachments: GIRAPH-306.patch > > > One of the biggest scalability challenges is getting Giraph to run reliably > on a large number of tasks (i.e. > 200). Several problems exist: > 1) If the connection fails after the initial connection was made, the job > will die. > 2) Requests must be completed exactly once. This is difficult to implement, > but required since we cannot have multiple retried requests succeed (i.e. a > vertex gets more messages than expected). > 3) Sometimes there are unresolved addresses, causing failure. > This patch addresses these issues by re-establishing failed connections and > keep tracking of every request sent to every worker. If the request fails or > passes a timeout, it will be resent. The server will keep track of requests > that succeeded to insure that the same request won't be processed more than > once. The structure for keeping track of the succeeded requests on the > server is efficient for handling increasing request ids (IncreasingBitSet). > For handling unresolved addresses, I added retry logic to keep trying to > resolve the problem. > This patch also adds several unit tests that use fault injection to simulate > a lost response or a closed channel exception on the server. It also has > unittests for IncreasingBitSet to insure it is working correctly and > efficiently. > This passes all unittests (including the new ones). Additionally, I have > some experience results as well. > Previously, I was unable to run reliably with more than 200 workers. With > this change I can reliably run 500+ workers. I also ran with 600 workers > successfully. This is a really big reliability win for us. > I can see the code working to do reconnections and re-issue requests when > necessary. It's very cool. > I.e. > 2012-08-18 00:16:52,109 INFO org.apache.giraph.comm.NettyClient: > checkAndFixChannel: Fixing disconnected channel to > xxx.xxx.xxx.xxx/xx.xx.xx.xx:30455, open = false, bound = false > 2012-08-18 00:16:52,111 INFO org.apache.giraph.comm.NettyClient: > checkAndFixChannel: Connected to xxx.xxx.xxx.xxx/xxx.xxx.xxx.xxx:30455! > 2012-08-18 00:16:52,123 INFO org.apache.giraph.comm.NettyClient: > checkAndFixChannel: Fixing disconnected channel to > xxx.xxx.xxx.xxx/xxx.xxx.xxx.xxx, open = false, bound = false > 2012-08-18 00:16:52,124 INFO org.apache.giraph.comm.NettyClient: > checkAndFixChannel: Connected to xxx.xxx.xxx.xxx/xxx.xxx.xxx.xxx:30117! -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira