Charles Connell created HBASE-29123:
---------------------------------------
Summary: CodecPool has performance issues
Key: HBASE-29123
URL: https://issues.apache.org/jira/browse/HBASE-29123
Project: HBase
Issue Type: Improvement
Reporter: Charles Connell
Assignee: Charles Connell
Attachments: borrow-decompressor.html, lease-counting.html,
return-decompressor.html
I look at many profile flamegraphs of my company's RegionServers. I sometimes
see memory allocation inside of {{org.apache.hadoop.io.compress.CodecPool}}
taking up roughly 1% of my CPU time. The point of a CodecPool is to avoid
allocating short-lived objects, so this is not good. Luckily, these allocations
can be avoided. Attached are three flamegraphs showing the allocations I'm
talking about.
I plan this ticket as the first of a series relating to decompression
performance. In the context of the overall series, it makes sense to fork
CodecPool out of hadoop-common and start a new copy of it in HBase. I'll do
that in this ticket and include my improvements.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
