Re: Re: Re: Memory operations on Sun/Oracle M class servers vs T class servers

very interessting. Thanks a lot.

Lothar
----Ursprüngliche Nachricht----
Von : contact_at_soocs.de
Datum : 17/12/2014 - 12:00 (GMT)
An : l.flatz_at_bluewin.ch
Cc : oracle-l_at_freelists.org
Betreff : Re: Re: Memory operations on Sun/Oracle M class servers vs T class servers

Hi Lothar,
my reply is a little bit OT regarding the original question, but maybe interesting for somebody.

>> I could work with IBM P7 recently. The acrchicture of the P7 is compareable in the sense that ist is also trying to squeeze more out of the core by
>> adding threads. Actually for demanding loads as highperforming databases the idea will not work out. Unforntunately it seems the trend is in favour
>> of more threads, as the P8 shows.

True, but luckily we can switch this SMT behavior dynamically with POWER architecture ( http://www-03.ibm.com/systems/resources/pwrsysperf_SMT4OnP7.pdf ). I have done some benchmarks on p770 with shared CPU pools and AIX 7.1 with SMT2 and SMT4. With SMT4 - the assigned entitlement could not be reached (all the CPUs were on fire) although there was enough free capacity in the shared CPU pool. However with SMT2 - the assigned entitlement was fully utilized (on each LPAR), even if all the CPUs were on fire.

Best Regards
Stefan Koehler

Oracle performance consultant and researcher Homepage: http://www.soocs.de
Twitter: _at_OracleSK

> "l.flatz_at_bluewin.ch" <l.flatz_at_bluewin.ch> hat am 17. Dezember 2014 um 11:37 geschrieben:
>
>
> Hi Finn,
>
> some years ago I discovered an issue on Solaris where the CPU ws slowed down by MMON. We did a similar test observing the performance of select from
> dual in a loop.
> The work around was shutting down automatic memory management.
> Thus for me it looks clearly like you have an issue here and it looks familiar.
>
> As for the CPU architecture I think your argument about the cores sharing a infrastructure makes sense. I could work with IBM P7 recently. The
> acrchicture of the P7 is compareable in the sense that ist is also trying to squeeze more out of the core by adding threads. Actually for demanding
> loads as highperforming databases the idea will not work out.
> IBM here suggests to switch off some core for more power on the database load (TurboCore:
> http://www.ibmsystemsmag.com/power/trends/ibmannouncements/power7_unveiled/TurboCore,-Active-Memory-Expansion-and-More/).
> Unforntunately it seems the trend is in favour of more threads, as the P8 shows.
>
> Regards
>
> Lothar

--
http://www.freelists.org/webpage/oracle-l



--
http://www.freelists.org/webpage/oracle-l