Re: AIX Large Pages and SMT4

From: Neil Chandler <neil_chandler_at_hotmail.com>
Date: Fri, 1 Nov 2019 16:39:47 +0000
Message-ID: <DB7PR10MB2090E47AD1EAC69DCDBE023785620_at_DB7PR10MB2090.EURPRD10.PROD.OUTLOOK.COM>



Cheers Stefan - I suppose the payoff between 64K/16M pages is the additional interaction needed with the Unix team when adjusting SGA's for (potentially) minimal benefit. Good point about the shape of the application. I've not seen any tangible benefit using SLOB with 16M pages, but I think I may need to adjust the amount of SGA being used during benchmarking to see any benefit fr the larger pages.

I suppose the kicker with SMT2/4/8 is the amount of CPU's generally active v the peak load (and will the peak load still be that peaky given the faster CPU's). Think I may have to revisit my benchmarking for the P9's that have just been delivered.

Good links - I'd not seen a couple of them previously in my googling.

Neil.



From: Stefan Koehler <contact_at_soocs.de> Sent: 25 October 2019 12:17
To: neil_chandler_at_hotmail.com <neil_chandler_at_hotmail.com>; Oracle-L_at_freelists.org <Oracle-L_at_freelists.org> Subject: Re: AIX Large Pages and SMT4

Hello Neil,

> Does anyone have any input on whether to use medium pages or large pages on AIX in relation to performance?

Since Oracle 11g medium (64k) page size is used by default - so the question would be more like any benefit from 64k to 16 MB page size. In general I would advise to use large pages as I have seen several performance improvements but have also seen cases where it has close to 0 impact - however these systems were not a real good candidate anyway because of highly I/O driven. It does not hurt but can improve :)

By the way even IBM states it in some of its official papers, e.g. http://www-03.ibm.com/support/techdocs/atsmastr.nsf/WebIndex/WP102171

> Does anyone have any input about using SMT2/4/8 usage - my slob testing has seen slightly better throughput with SMT4 over SMT2 (SMT4 vs SMT8 was the same)

My advice: Check your production workload and set SMT to 1 if #num_of_session_on_cpu =< #amount_of_cores. You get the best single thread performance with no SMT at all ( https://www.anandtech.com/show/10435/assessing-ibms-power8-part-1/4 & https://www.ibm.com/developerworks/community/blogs/aixpert/entry/local_near_far_memory_part_3_scheduling_processes_to_smt_virtual_processors130 ) - Oracle is mostly a single thread application and so don't waste possibly available CPU instructions :-)

As soon as you have more processes on CPU than cores you benefit from higher SMT levels - however AFAIK the dispatching has slightly changed with Power 9 and I never did any benchmarks with that platform in this area.

Best Regards
Stefan Koehler

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

> Neil Chandler <neil_chandler_at_hotmail.com> hat am 24. Oktober 2019 um 10:48 geschrieben:
>
>
> Hi,
>
> A couple of Oracle (11.2.0.4 and 19) on AIX (7.2) questions:
>
> Does anyone have any input on whether to use medium pages or large pages on AIX in relation to performance? I'm primarily concerned about performance - especially when the server is stressed.
>
> Does anyone have any input about using SMT2/4/8 usage - my slob testing has seen slightly better throughput with SMT4 over SMT2 (SMT4 vs SMT8 was the same) but my testing has been limited to this point and is far from conclusive.
>
> Neil Chandler.

--
http://www.freelists.org/webpage/oracle-l
Received on Fri Nov 01 2019 - 17:39:47 CET

Original text of this message