On 7/8/13 7:32 PM, Mircea Markus wrote:
Indeed adding the DONT_BUNDLE flag back and using the new bundler
yielded the same performance as the old setup.
That's bad, we should look into it ! I get exactly the same or even
better performance with the new bundler and DONT_BUNDLE removed.
Interesting enough, using the DONT_BUNDLE + old bundler made
performance not look good.
Radim can provide the exact numbers.
That's completely counter to what I've measured; and it should be
exactly the other way round !
On 25 Jun 2013, at 16:37, Bela Ban <bban(a)redhat.com> wrote:
> So the new bundler is always *faster* than the old one. And ISPN-3221
> seems to fix the degradation. Do I still need to look into this ?
>
> On 6/13/13 12:43 PM, Radim Vansa wrote:
>>
>> | > Seems that it's somewhat better (6050 writes/s) but not at the
previous
>> | > level.
>> | what about the reads?
>>
>> After Pedro's question I have tried to run ISPN-3221 fix with both old an new
bundler (in configuration), and the result pretty surprises me as there should not be much
of a difference (as all messages that are part of the test should be OOB =>
DONT_BUNDLE).
>>
>> 113842c8cf91cbb5a1bbd26e05fab7024fdec081 (last OK build)
>> new bundler: 1.5M reads/s, 7800 writes/s
>> old bundler: 1.3M reads/s, 4300 writes/s
>>
>> 73da108cdcf9db4f3edbcd6dbda6938d6e45d148 (removing DONT_BUNDLE)
>> new bundler: 1.2M reads/s, 4800 writes/s
>> old bundler: 150k reads/s, 250 writes/s
>>
>> ISPN-3221 fix (Pedro's branch)
>> new bundler: 1.5M reads/s, 6050 writes/s
>> old bundler: 1.3M reads/s, 3100 writes/s
>>
>> ISPN-3221 cherry-picked onto 73da108cdcf9db4f3edbcd6dbda6938d6e45d148
>> new bundler: 1.5M reads/s, 7800 writes/s
>> old bundler: 1.3M reads/s, 4300 writes/s
Bela Ban, JGroups lead (
http://www.jgroups.org)