Looks you are using apache-hadoop 2.6.
a) Are there any promotion failures are there..? can be checked using following. -XX:+PrintGCDetails -XX:+PrintPromotionFailure -XX:PrintFLSStatistics=1 b) SurvivorRatio is configured..? can you give all the GC params configured.? Even GC logs.? the option -XX:+PrintTenuringDistribution to show the threshold and ages of the objects in the new generation. It is useful for observing the lifetime distribution of an application. Periodic heapdump could have helped. Did you crossed checked the following also. i) NN and JN have dedicated disk's to for writing edit log transactions. ii) " dfs.namenode.accesstime.precision" should n't be set low value,NN will update last accessed time for each file iii) Group look up isn't taking more time,as nn will check for every request. iv) NN log for Excessive Spew ( there are some jira's which are fixed later to 2.6.0) like decommission nodes,... v) Any recursive calls are very frequently? -----Original Message----- From: Lin,Yiqun(vip.com) [mailto:yiqun01....@vipshop.com] Sent: Wednesday, September 26, 2018 10:26 AM To: Wei-Chiu Chuang <weic...@cloudera.com> Cc: Hdfs-dev <hdfs-dev@hadoop.apache.org> Subject: 答复: NN run progressively slower Hi Wei-Chiu, At the beginning, we have noted HDFS-9260 which changing the structure of the stored blocks. With multiple adding/removing under this structure, there will be a performance degradation. But we think this will reach a stable state, and HDFS-9260 isn’t the root cause from our point of view. >Did you observe clients failing to close file due to insufficient number of >block replicas? Did NN fail over? No failing close file warning in client. And no NN fail over happened. >Did you have gc logging enabled? Any chance to take a heap dump and analyze >what's in there? We have enabled gc logging but haven’t take the analysis for the NN heap dump. From the gc log, we find average time of ygc around 0.02s but total ygc time within one day are increasing day by day. >There were quite some NN scalability and GC improvements between CDH5.5 ~ >CDH5.8 time frame. We have customers at/beyond your scale in your version but >I don't think I've heard similar symptoms. Thanks for sharing this. Thanks Yiqun 发件人: Wei-Chiu Chuang [mailto:weic...@cloudera.com] 发送时间: 2018年9月26日 11:14 收件人: 林意群[运营中心] 抄送: Hdfs-dev 主题: Re: NN run progressively slower Yiqun, Is this related to HDFS-9260? Note that HDFS-9260 was backported since CDH5.7 and above. I'm interested to learn more. Did you observe clients failing to close file due to insufficient number of block replicas? Did NN fail over? Did you have gc logging enabled? Any chance to take a heap dump and analyze what's in there? There were quite some NN scalability and GC improvements between CDH5.5 ~ CDH5.8 time frame. We have customers at/beyond your scale in your version but I don't think I've heard similar symptoms. Regards On Tue, Sep 25, 2018 at 2:04 AM Lin,Yiqun(vip.com<http://vip.com>) <yiqun01....@vipshop.com<mailto:yiqun01....@vipshop.com>> wrote: Hi hdfs developers: We meet a bad problem after rolling upgrade our hadoop version from 2.5.0-cdh5.3.2 to 2.6.0-cdh5.13.1. The problem is that we find NN running slow periodically (around a week). Concretely to say, For example, we startup NN on Monday, it will run fast. But time coming to Weekends, our cluster will become very slow. In the beginning, we thought maybe some FSN lock caused by this. And we did some improvements for this, e.g. configurable the remove block interval, print FSN lock elapsed time. After this, the problem still exists, :(. So we suspect this maybe not a hdfs rpc problem. Finally we find a related phenomenon: every time NN runs slow, its old gen reaches a high value, around 100GB. Actually, NN total metadata size is just around 40GB in our clsuter. So for the temporary solution, we reduce the heap space and trigger full gc frequently. Now it looks better than before but we haven’t found the root cause of this. Not so sure if this is a jvm tuning problem or a hdfs bug? Anyone who has met the similar problem in this version? Why the NN old gen space greatly increased? Some information of our env: JDK1.8 500+ Nodes, 150 million blocks, around 40GB metadata size will be used. Appreciate if anyone who can share your comments. Thanks Yiqun. 本电子邮件可能为保密文件。如果阁下非电子邮件所指定之收件人,谨请立即通知本人。敬请阁下不要使用、保存、复印、打印、散布本电子邮件及其内容,或将其用于其他任何目的或向任何人披露。谢谢您的合作! This communication is intended only for the addressee(s) and may contain information that is privileged and confidential. You are hereby notified that, if you are not an intended recipient listed above, or an authorized employee or agent of an addressee of this communication responsible for delivering e-mail messages to an intended recipient, any dissemination, distribution or reproduction of this communication (including any attachments hereto) is strictly prohibited. If you have received this communication in error, please notify us immediately by a reply e-mail addressed to the sender and permanently delete the original e-mail communication and any attachments from all storage devices without making or otherwise retaining a copy. -- A very happy Clouderan 本电子邮件可能为保密文件。如果阁下非电子邮件所指定之收件人,谨请立即通知本人。敬请阁下不要使用、保存、复印、打印、散布本电子邮件及其内容,或将其用于其他任何目的或向任何人披露。谢谢您的合作! This communication is intended only for the addressee(s) and may contain information that is privileged and confidential. You are hereby notified that, if you are not an intended recipient listed above, or an authorized employee or agent of an addressee of this communication responsible for delivering e-mail messages to an intended recipient, any dissemination, distribution or reproduction of this communication (including any attachments hereto) is strictly prohibited. If you have received this communication in error, please notify us immediately by a reply e-mail addressed to the sender and permanently delete the original e-mail communication and any attachments from all storage devices without making or otherwise retaining a copy. --------------------------------------------------------------------- To unsubscribe, e-mail: hdfs-dev-unsubscr...@hadoop.apache.org For additional commands, e-mail: hdfs-dev-h...@hadoop.apache.org