[infinispan-dev] performance regression with replication

[Mircea Markus]

Thanks Radim,

My assumption was only partially correct as we still send an unnecessary TxCompletionNotificationCommand in this scenario.
I've made a patch to remove the TxCompletionNotificationCommand, can you please give it a go: https://github.com/infinispan/infinispan/pull/1901



On 14 Jun 2013, at 08:20, Radim Vansa <rvansa at redhat.com> wrote:

> Hi Mircea,
> 
> I am afraid that your assumption is wrong - tx cache with 1pc autocommit, accessed non-transactionally has about 1.1M - 1.2M reads/s and write performance varies between 3600 and 6900 writes/s (specifically, in one test it was 3600, 4300, 4500 and 6900). That's surprising, I've checked that the consistent hash is spread evenly.
> I am talking about new bundler and all changes (ISPN-2848 + ISPN-3221, ISPN-2772 + ISPN-3196) merged in - I've checked the performance on both pruivo's branch and my t_regression with reordered commits.
> 
> Do you still need the JProfiler output with/without ISPN-2772?
> 
> Radim
> 
> ----- Original Message -----
> | From: "Mircea Markus" <mmarkus at redhat.com>
> | To: "infinispan -Dev List" <infinispan-dev at lists.jboss.org>
> | Sent: Thursday, June 13, 2013 9:30:56 PM
> | Subject: [infinispan-dev] performance regression with replication
> | 
> | Hi Radim,
> | 
> | The performance regression introduced with ISPN-2772 is caused by the fact
> | that the new replication mechanism in 5.3 is now using lock delegation[1].
> | This allows for multiple writes to happen concurrently: both guarantees
> | consistency when there are no failures and avoids deadlocks. Support for
> | concurrent writes comes at the cost of an additional (delegating) RPC per
> | request - hence the drop in performance.
> | 
> | So we're not really comparing apples with apples here. Also a transactional
> | cache configured with 1PcForAutoCommitTransaction=true, which is
> | functionally equivalent with 5.2.x's replication, has the same number of
> | RPCs and consequently I expect pretty similar performance with the repl from
> | 5.2.x. Can you please compare these two? If I'm right then we can recommend
> | all the users that don't need concurrent writes to use tx +
> | 1PcForAutoCommitTransaction=true. If I'm not we should provide a backward
> | compatible layer for the sake of performance.
> | 
> | [1]
> | https://docs.jboss.org/author/display/ISPN/Locking+and+Concurrency#LockingandConcurrency-Nontransactionalcachesandconcurrentupdates%26nbsp%3B
> | 
> | 
> | Cheers,
> | --
> | Mircea Markus
> | Infinispan lead (www.infinispan.org)
> | 
> | _______________________________________________
> | infinispan-dev mailing list
> | infinispan-dev at lists.jboss.org
> | https://lists.jboss.org/mailman/listinfo/infinispan-dev
> | 
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/infinispan-dev

Cheers,
-- 
Mircea Markus
Infinispan lead (www.infinispan.org)