On 26 Nov 2010, at 14:06, Eduardo Martins wrote:

Currently JBoss Cache (along with JTA) is what is holding off
Mobicents performance improvements, a profiling session will put
everything related with JBC interaction as top hotspots. Note that
currently performance is not bad at all, we are at same level as
competition but in this business, and unlike Java EE, latency and
throughput is all that matters.

Tell me more about this indirection/references layer, since I believe
no locks on child insert/removals will be a must. Also note that we
will need to keep a transaction isolation similar to what JBoss Cache
provides now, so 2 parallel child inserts/removals must be handled
correctly at JTA level too, dunno if it matters at all but just want
to define what are the requirements.

Well, step 1 should be for you to have a look at the srcs of the Tree module and see how it breaks up tree structured data into flat data. If this is what you would have done manually anyway, then it's already done for you. :-)

https://github.com/infinispan/infinispan/tree/master/tree/

-- Eduardo
..............................................
http://emmartins.blogspot.com
http://redhat.com/solutions/telco

On Fri, Nov 26, 2010 at 1:54 PM, Manik Surtani <manik@jboss.org> wrote:

On 26 Nov 2010, at 13:38, Eduardo Martins wrote:

You mean we will get 5-10% if implementing the parent, child relations
ourselves, where needed?

Maybe. Since some of the relationships could be denormalized/flattened, etc.

That would for sure be worth the data model
rework.

Depends. How much of your time is spent in the data grid? 5%? So 5% x 5% = overall 0.25% gain? :-)

The child relation would be done using a cache map entry with
a set/array value? If yes wouldn't that lock the parent on each child
insert/removal?

Yes. Or you could use a layer of indirection/references which will then not lock the hierarchy.

-- Eduardo
..............................................
http://emmartins.blogspot.com
http://redhat.com/solutions/telco

On Fri, Nov 26, 2010 at 1:05 PM, Manik Surtani <manik@jboss.org> wrote:

On 25 Nov 2010, at 03:24, Eduardo Martins wrote:

Great, finally an easy migration from JBoss Cache.

Just to know our options in Mobicents future, are there any cons on
using this API? Is there a reason to change a data model/schema coming
from JBoss Cache, that really takes advantage of node relations
parent/child? For instance it's very common to retrieve the set of
childs for a node, another common usage is the removal for a complete
subtree. We have time to rework this data model if it's not the best
usage, we want it as fast as possible, both local and cluster/cloud
modes.

There is nothing wrong with using the Tree API, except that it doesn't perform as well as the flat API. Now don't let this scare you off - in terms of performance, we're talking 5 ~ 10% and if the API works better for you, then use it. But you should benchmark for yourself and see whether you need to extract this extra performance in exchange for reworking your data model. YMMV. :-)

Cheers
Manik

--
Manik Surtani
manik@jboss.org
Lead, Infinispan
Lead, JBoss Cache
http://www.infinispan.org
http://www.jbosscache.org

_______________________________________________
infinispan-dev mailing list
infinispan-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev

_______________________________________________
infinispan-dev mailing list
infinispan-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev

--
Manik Surtani
manik@jboss.org
Lead, Infinispan
Lead, JBoss Cache
http://www.infinispan.org
http://www.jbosscache.org

_______________________________________________
infinispan-dev mailing list
infinispan-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev

_______________________________________________
infinispan-dev mailing list
infinispan-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev