[infinispan-dev] Performance improvements, more...

[Bela Ban]

On 1/19/12 12:03 PM, Manik Surtani wrote:
> All looking very good, people.  Bela, feel like putting Table into NAKACK and 3.0.3?  ;)

Definitely not in NAKACK, it will be in NAKACK2 in 3.1. Once this has 
been running in production for a year or so, we can rename it to NAKACK.

Remember, NakReceiverWindow might not be the fastest class around, but 
it's very stable and *correct*, and has been around for 12+ years... :-)

If my experiments with NAKACK2 in 3.1 prove successful and I'm done 
sooner than expected, then, yes, I can backport it to 3.0.3 (or 3.0.4). 
I'd prefer to have Sanne run a test with the experimental 3.1 first 
though, so we can see if this makes a difference at all...




> On 19 Jan 2012, at 16:29, Bela Ban wrote:
>
>>
>>
>> On 1/19/12 11:45 AM, Sanne Grinovero wrote:
>>> On 19 January 2012 09:59, Bela Ban<bban at redhat.com>   wrote:
>>>> It would be interesting to see the numbers with bbc128, which makes
>>>> sending a bit faster. I'd expect to see more writes and less reads,
>>>> compared to their relative numbers.
>>>
>>> Ok, done. This is the same Infinispan build, but using JGroups bbc128:
>>>
>>> Done 880,969,860 transactional operations in 24.71 minutes using 5.1.0-SNAPSHOT
>>>    875,033,689 reads and 5,936,171 writes
>>>    Reads / second: 590,216
>>>    Writes/ second: 4,003
>>
>>
>> OK, thanks. Not as dramatic though as the change in 23a031e...
>>
>>
>>> Looks like a bit slower - confirming the figures I had two days ago.
>>> Anyway my purpose with the comparison was just to proof the latest
>>> patches in Infinispan where going in the correct direction, so I'm
>>> intentionally not changing JGroups versions yet.
>>>
>>>> BTW: I'm done with my implementation of Table, and the numbers look
>>>> really impressive ! It is about the same as RingBuffer for smaller
>>>> insertions (5 million), but for 50 million the number stays about the
>>>> same (insertions and removals per second). For smaller numbers, Table is
>>>> ca 4 times *faster* than NakReceiverWindow.
>>>>
>>>> I still want to add more tests for Table (copy and convert the ones for
>>>> RingBuffer), and then switch NAKACK2 over from RingBuffer to Table. I'm
>>>> very curious to see the perf numbers after that change !
>>>>
>>>> Next comes passing up of entire bundles, this should also make a big
>>>> difference !
>>>> Exiting times, cheers !
>>>
>>> If you commit it on an experimental branch, I'll give it a preview run ..
>>
>> The branch is JGRP-1396-2, the class is Table. There is a stress test
>> called TableStressTest (you can compare it to
>> NakReceiverWindowStressTest and RingBufferStressTest).
>>
>>
>> --
>> Bela Ban
>> Lead JGroups (http://www.jgroups.org)
>> JBoss / Red Hat
>> _______________________________________________
>> infinispan-dev mailing list
>> infinispan-dev at lists.jboss.org
>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>
> --
> Manik Surtani
> manik at jboss.org
> twitter.com/maniksurtani
>
> Lead, Infinispan
> http://www.infinispan.org
>
>
>
>
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/infinispan-dev

-- 
Bela Ban
Lead JGroups (http://www.jgroups.org)
JBoss / Red Hat