[
https://issues.apache.org/jira/browse/CASSANDRA-7542?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14073445#comment-14073445
]
Benedict commented on CASSANDRA-7542:
-------------------------------------
I decided to have a quick crack at this to see if I could get an initial patch
done for you to take for a spin. It's available
[here|https://github.com/belliottsmith/cassandra/tree/7542-cascontend]
I find on my local machine that it speeds up the paxos dtest by about 20%,
which may be a little conservative for improvement as it takes a little while
for it to get an idea of how long a paxos round takes. There are some
improvements that should probably be made before rolling it out generally, such
as tracking latency for each token range independently, and we should probably
track the level of recent contention so that we can exponentially backoff if
we're too aggressive (this might permit us to be a little _more_ aggressive in
the typical case).
On my box, it's worth noting this patch brings the average wait time due to
contention down to around 15ms, instead of 50ms. Since this is more than a 20%
decline, there is probably some more tuning to be done besides to get improved
throughput.
This only explores two of the potential ideas: reducing intra-node competition,
and reducing sleep interval.
> Reduce CAS contention
> ---------------------
>
> Key: CASSANDRA-7542
> URL: https://issues.apache.org/jira/browse/CASSANDRA-7542
> Project: Cassandra
> Issue Type: Improvement
> Reporter: sankalp kohli
> Assignee: Benedict
> Fix For: 2.0.10
>
>
> CAS updates on same CQL partition can lead to heavy contention inside C*. I
> am looking for simple ways(no algorithmic changes) to reduce contention as
> the penalty of it is high in terms of latency, specially for reads.
> We can put some sort of synchronization on CQL partition at StorageProxy
> level. This will reduce contention at least for all requests landing on one
> box for same partition.
> Here is an example of why it will help:
> 1) Say 1 write and 2 read CAS requests for the same partition key is send to
> C* in parallel.
> 2) Since client is token-aware, it sends these 3 request to the same C*
> instance A. (Lets assume that all 3 requests goto same instance A)
> 3) In this C* instance A, all 3 CAS requests will contend with each other in
> Paxos. (This is bad)
> To improve contention in 3), what I am proposing is to add a lock on
> partition key similar to what we do in PaxosState.java to serialize these 3
> requests. This will remove the contention and improve performance as these 3
> requests will not collide with each other.
> Another improvement we can do in client is to pick a deterministic live
> replica for a given partition doing CAS.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
