[
https://issues.apache.org/jira/browse/CASSANDRA-15213?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17015512#comment-17015512
]
Benedict Elliott Smith commented on CASSANDRA-15213:
----------------------------------------------------
fwiw, wrt bucket distribution, we should be able to do something as simple as
multiplying by the smallest prime larger than two cache-lines (when multiplied
by 8), that doesn't also divide the number of buckets. i.e., 17 (with
verification it doesn't divide a custom number of buckets, and choosing another
prime if it does)
i.e.
{code}
stripedIndex(index, stripe) = ((index * prime) % (bucketOffsets.length + 1)) +
((bucketOffsets.length + 1) * stripe)
or
stripedIndex(index, stripe) = (((index * nStripes + stripe) * prime) %
buckets.length())
{code}
> DecayingEstimatedHistogramReservoir Inefficiencies
> --------------------------------------------------
>
> Key: CASSANDRA-15213
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15213
> Project: Cassandra
> Issue Type: Bug
> Components: Observability/Metrics
> Reporter: Benedict Elliott Smith
> Assignee: Jordan West
> Priority: Normal
> Fix For: 4.0-beta
>
>
> * {{LongAdder}} introduced to trunk consumes 9MiB of heap without user
> schemas, and this will grow significantly under contention and user schemas
> with many tables. This is because {{LongAdder}} is a very heavy class
> designed for single contended values.
> ** This can likely be improved significantly, without significant loss of
> performance in the contended case, by simply increasing the size of our
> primitive backing array and providing multiple buckets, with each thread
> picking a bucket to increment, or simply multiple backing arrays. Probably a
> better way still to do this would be to introduce some competition detection
> to the update, much like {{LongAdder}} utilises, that increases the number of
> backing arrays under competition.
> ** To save memory this approach could partition the space into chunks that
> are likely to be updated together, so that we do not need to duplicate the
> entire array under competition.
> * Similarly, binary search is costly and a measurable cost as a share of the
> new networking work (without filtering it was > 10% of the CPU used overall).
> We can compute an approximation floor(log2 n / log2 1.2) extremely cheaply,
> to save the random memory access costs.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
