You can try turning on verbose garbage collection logs and see if the slow times correspond to a GC pause. Cloudera has a series of blog posts regarding GC pauses in HBase and how to avoid them: http://www.cloudera.com/blog/2011/02/avoiding-full-gcs-in-hbase-with-memstore-local-allocation-buffers-part-1/
----- Original Message ----- From: "Assarsson, Emil" <[email protected]> To: "'[email protected]'" <[email protected]> Cc: Sent: Monday, February 6, 2012 9:15 AM Subject: randomWrite tests gives random results Hi, I'm tryng to optimize a hbase cluster (on hdfs) with the test randomWrite. I have 7 nodes: 1 zookeeper/name/hbase-master/jobtracker and 6 region/data/tasktrackers. Each with 1 disk, 16G memory, 2 x 4 cores. I know that I really should have more disks but for the time being I'm trying to do the best with what I have. I have configured tasktrackers to run 1 map/1 red on each host. The problem I'm seeing is that I get very varying results spanning from 16sec/100000inserts to 240sec/100000inserts. Currently I'm using a 10G heapsize on hbase and 3G heapsize on hdfs. How do I find out what makes it this random? I think I should be able to get around 22sec/100000inserts. Best regards Emil Assarsson Sony Ericsson Mobile Communications AB "The information in this email, and attachment(s) thereto, is strictly confidential and may be legally privileged. It is intended solely for the named recipient(s), and access to this e-mail, or any attachment(s) thereto, by anyone else is unauthorized. Violations hereof may result in legal actions. Any attachment(s) to this e-mail has been checked for viruses, but please rely on your own virus-checker and procedures. If you contact us by e-mail, we will store your name and address to facilitate communications in the matter concerned. If you do not consent to us storing your name and address for above stated purpose, please notify the sender promptly. Also, if you are not the intended recipient please inform the sender by replying to this transmission, and delete the e-mail, its attachment(s), and any copies of it without, disclosing it."
