[ https://issues.apache.org/jira/browse/HBASE-5347?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13215085#comment-13215085 ]
Prakash Khemani commented on HBASE-5347: ---------------------------------------- Lars, you are right. I have been trying to induce a Full GC but without any success. (I can induce a full GC if I artificially hold some key-values in queue and force them to be tenured.) On 89-fb, my test-case is doing random increments on a space of slightly more than 40GB worth of Key-value data. The heap is set to 36GB. The LRU cache has a high and low watermark of .98 and .85 percents. The region server spawns 1000 threads that continuously do the increments. The eviction thread manages to keep the block-cache at about 85% always. Cache-on-write is turned on to induce more cache churn. All the 12 disks are close to 100% read pegged. GC takes 60% of the CPU (sum of user times in 1000 lines of gc log / (elapsed time * #cpus)). Compactions that get started never complete while the load is on. I guess I have to change the dynamics of the test case to induce GC pauses. On 2/22/12 11:35 PM, "Todd Lipcon (Commented) (JIRA)" <j...@apache.org> wrote: > GC free memory management in Level-1 Block Cache > ------------------------------------------------ > > Key: HBASE-5347 > URL: https://issues.apache.org/jira/browse/HBASE-5347 > Project: HBase > Issue Type: Improvement > Reporter: Prakash Khemani > Assignee: Prakash Khemani > Attachments: D1635.5.patch > > > On eviction of a block from the block-cache, instead of waiting for the > garbage collecter to reuse its memory, reuse the block right away. > This will require us to keep reference counts on the HFile blocks. Once we > have the reference counts in place we can do our own simple > blocks-out-of-slab allocation for the block-cache. > This will help us with > * reducing gc pressure, especially in the old generation > * making it possible to have non-java-heap memory backing the HFile blocks -- 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