[
https://issues.apache.org/jira/browse/CASSANDRA-16072?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17186401#comment-17186401
]
Benjamin Lerer edited comment on CASSANDRA-16072 at 8/28/20, 9:14 AM:
----------------------------------------------------------------------
I might be missing something but {{HintsBuffer.allocateBytes}} logic seems
wrong.
{code}
private int allocateBytes(int totalSize)
{
int prev = position.getAndAdd(totalSize);
if (prev == CLOSED) // the slab has been 'closed'
return CLOSED;
if ((prev + totalSize) > slab.capacity())
{
position.set(CLOSED); // mark the slab as no longer allocating if
we've exceeded its capacity
return CLOSED;
}
return prev;
}
{code}
As the following scenario will not produce the expected out come:
# Thread A calls {{allocateBytes}} and exit with the return value {{CLOSED}}
and {{position == CLOSED}}
# Thead B calls {{allocateBytes}} with {{totalSize == size_B}} and exit with
the return value {{CLOSED}} and {{position == size_B -1}} as {{CLOSED == -1}}
in practice.
# Thread C calls {{allocateBytes}} with {{totalSize == size_C}} and exit with
the return value {{size_B - 1}} and and {{position == size_B + size_C -1}}
was (Author: blerer):
I might be missing something but {{HintsBuffer.allocateBytes}} logic seems
wrong.
{code}
private int allocateBytes(int totalSize)
{
int prev = position.getAndAdd(totalSize);
if (prev == CLOSED) // the slab has been 'closed'
return CLOSED;
if ((prev + totalSize) > slab.capacity())
{
position.set(CLOSED); // mark the slab as no longer allocating if
we've exceeded its capacity
return CLOSED;
}
return prev;
}
{code}
As the following scenario will not produce the expected out come:
# Thread A calls {{allocateBytes}} and exit with the return value {{CLOSED}}
and {{position == CLOSED}}
# Thead B calls {{allocateBytes}} with {{totalSize == size_B}} and exit with
the return value {{CLOSED}} and and {{position == size_B -1}} as {{CLOSED ==
-1}} in practice.
# Thread C calls {{allocateBytes}} with {{totalSize == size_C}} and exit with
the return value {{size_B - 1}} and and {{position == size_B + size_C -1}}
> Reduce thread contention in CommitLogSegment and HintsBuffer by rewriting CAS
> loops to atomic adds
> --------------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-16072
> URL: https://issues.apache.org/jira/browse/CASSANDRA-16072
> Project: Cassandra
> Issue Type: Improvement
> Components: Consistency/Hints, Local/Commit Log
> Reporter: Michael Semb Wever
> Assignee: Michael Semb Wever
> Priority: Normal
> Fix For: 3.11.x, 4.0-beta
>
>
> Follow up to CASSANDRA-15922
> Both CommitLogSegment and HintsBuffer use AtomicIntegers for the current
> offset when allocating. Like in CASSANDRA\-15922 the loops on
> {{.compareAndSet(..)}} can be replaced with atomic adds using the {{.
> getAndAdd(..)}} method.
> In highly contended environments the CAS failures can be high, starving
> writes in a running Cassandra node. On the same cluster CASSANDRA\-15922 was
> found, after CASSANDRA\-15922's fix was deployed, there was still problems
> around commit log flushing and hints. No flamegraph was collected that
> demonstrated the thread contention as clearly as was found in
> CASSANDRA\-15922, but the performance fix proposed here hopefully is obvious
> enough.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
