Becker Ewing created HBASE-29099:
------------------------------------
Summary: Avoid expensive BlockCacheKey.toString() calls when
tracing is disabled
Key: HBASE-29099
URL: https://issues.apache.org/jira/browse/HBASE-29099
Project: HBase
Issue Type: Improvement
Components: Performance
Affects Versions: 2.5.11, 2.6.1, 3.0.0-beta-1, 4.0.0-alpha-1
Reporter: Becker Ewing
Assignee: Becker Ewing
Attachments: high-block-cache-key-to-string-alloc-profile.html
Currently, when reading a block from the BlockCache, [we always calculate
BlockCacheKey.toString()|https://github.com/apache/hbase/blob/b89c8259c5726395c9ae3a14919bd192252ca517/hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileReaderImpl.java#L1290-L1296]
so that we can add it to the Tracing Span. In a region server running
extremely hot (pinned at 100% CPU utilization for many hours due to a reverse
scanning hotspot), I noticed that in an allocation profile—see attached—of the
region server that almost 9% of all allocations were to calculate
BlockCacheKey.toString() which I found to be rather unusual.
Given that BlockCacheKey.toString() is only used to enhance traceability of
region servers, it feels like we should be able to skip calling this expensive
function is we can determine that tracing is disabled.
Note: this function was already identified as inefficient and had it's
performance optimized as part of HBASE-19684. I'm not certain whether we can do
much better than it's current implementation.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
