Dmitry Konstantinov created CASSANDRA-21754:
-----------------------------------------------
Summary: Reduce contention on MemtablePool.SubPool#allocated
Key: CASSANDRA-21754
URL: https://issues.apache.org/jira/browse/CASSANDRA-21754
Project: Apache Cassandra
Issue Type: Improvement
Components: Local/Memtable
Reporter: Dmitry Konstantinov
Assignee: Dmitry Konstantinov
During a write to a memtable we increase allocated counter within
MemtablePool.SubPool and check if it is crossed a limit to trigger a clean
(flush). We have multiple threads which update and read the value via
AtomicLongFieldUpdater. For a high write rate the field can become a contention
point + we may have false sharing for other fields within the same
MemtablePool.SubPool object. MemtablePool.SubPool is common across all tables
on a node.
To reduce the contention we can use the fact that we do not need to read the
allocated value every time, it is only needed to check if we've crossed a
threshold for the next flush.
So, we can apply a sharding approach for this value: we can split the value to
several sub-values which are updated by different threads independently
(similar to LongAdder). There are several ways to deal with threshold tracking,
for example we can have a smaller threshold for each shard value or do some
reservation from a common value, etc (a subject to evaluate the best one).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
