That really depends on your consumer application's memory allocation patterns. If it is a thin wrapper over a Kafka consumer, I would imagine you can get away with using CMS for the tenured generation and parallel collector for the new generation with a small heap like 1gb or so.
Thanks, Neha On Monday, March 25, 2013, Yonghui Zhao wrote: > Any suggestion on consumer side? > 在 2013-3-25 下午9:49，"Neha Narkhede" <neha.narkh...@gmail.com <javascript:;> > >写道： > > > For Kafka 0.7 in production at Linkedin, we use a heap of size 3G, new > gen > > 256 MB, CMS collector with occupancy of 70%. > > > > Thanks, > > Neha > > > > On Sunday, March 24, 2013, Yonghui Zhao wrote: > > > > > Hi Jun, > > > > > > I used kafka-server-start.sh to start kafka, there is only one jvm > > setting > > > "-Xmx512M“ > > > > > > Do you have some recommend GC setting? Usually our sever has 32GB or > > 64GB > > > RAM. > > > > > > 2013/3/22 Jun Rao <jun...@gmail.com> > > > > > > > A typical reason for many rebalancing is the consumer side GC. If so, > > you > > > > will see logs in the consume saying sth like "expired session" for > ZK. > > > > Occasional rebalances are fine. Too many rebalances can slow down the > > > > consumption and you will need to tune your GC setting. > > > > > > > > Thanks, > > > > > > > > Jun > > > > > > > > On Thu, Mar 21, 2013 at 11:07 PM, Yonghui Zhao < > zhaoyong...@gmail.com > > > > >wrote: > > > > > > > > > Yes, before consumer exception: > > > > > > > > > > 2013/03/21 12:07:17.909 INFO [ZookeeperConsumerConnector] [] > > > > > 0_lg-mc-db01.bj-1363784482043-f98c7868 *end rebalancing > > > > > consumer*0_lg-mc-db01.bj-1363784482043-f98c7868 try #0 > > > > > 2013/03/21 12:07:17.911 INFO [ZookeeperConsumerConnector] [] > > > > > 0_lg-mc-db01.bj-1363784482043-f98c7868 *begin rebalancing > > > > > consumer*0_lg-mc-db01.bj-1363784482043-f98c7868 try #0 > > > > > 2013/03/21 12:07:17.934 INFO [FetcherRunnable] [] FetchRunnable-0 > > start > > > > > fetching topic: sms part: 0 offset: 43667888259 from > 127.0.0.1:9093 > > > > > 2013/03/21 12:07:17.940 INFO [SimpleConsumer] [] Reconnect in > > > multifetch > > > > > due to socket error: > > > > > java.nio.channels.*ClosedByInterruptException* > > > > > at java.nio.channels.spi.*AbstractInterruptibleChannel* > > > > > .end(AbstractInterruptibleChannel.java:201) > > > > > > > > > > > > > > > 2013/03/21 12:07:17.978 INFO [ZookeeperConsumerConnector] [] > > > > > 0_lg-mc-db01.bj-1363784482043-f98c7868 *end rebalancing > > > > > consumer*0_lg-mc-db01.bj-1363784482043-f98c7868 try #0 > > > > > 2013/03/21 12:07:18.004 INFO [FetcherRunnable] [] FetchRunnable-0 > > start > > > > > fetching topic: sms part: 0 offset: 43667888259 from > 127.0.0.1:9093 > > > > > 2013/03/21 12:07:18.066 INFO [ZookeeperConsumerConnector] [] > > > > > 0_lg-mc-db01.bj-1363784482043-f98c7868 *begin rebalancing consume*r > > > > > 0_lg-mc-db01.bj-1363784482043-f98c7868 try #0 > > > > > 2013/03/21 12:07:18.176 INFO [SimpleConsumer] [] Reconnect in > > > multifetch > > > > > due to socket error: > > > > > java.nio.channels.*ClosedByInterruptException* > > > > > at java.nio.channels.spi.*AbstractInterruptibleChannel* > > > > > .end(AbstractInterruptibleChannel.java:201) > > > > > > > > > > > > > > > So you think it is normal? How can we avoid this exception? > > > > > > > > > > I used 4 partitions in kafka, use only 1 partition？ > > > > > > > > > > > > > > > > > > > > 2013/3/22 Jun Rao <jun...@gmail.com> > > > > > > > > > > > Do you see any rebalances in the consumer? Each rebalance will > > > > interrupt > > > > > > existing fetcher threads first. > > > > > > > > > > > > Thanks, > > > > > > > > > > > > Jun > > > > > > > > > > > > On Thu, Mar 21, 2013 at 9:40 PM, Yonghui Zhao < > > zhaoyong...@gmail.com >
