RE: ASMM - resizing triggers/thresholds

From: Mark W. Farnham <mwf_at_rsiz.com>
Date: Mon, 4 Feb 2019 10:05:22 -0500
Message-ID: <10eb01d4bc9b$0d46bf00$27d43d00$_at_rsiz.com>

Slightly askew, old advice memory triggered by JL's mention of pinned:

Survey what you chronically use, most especially things that you use intermittently but which fall on the slope of the "worth it to keep in memory to avoid the reload" and put them in your warm start procedure for any instances serviced by listeners that use them (either primary and secondary or everything, depending on your environment).

Consider strongly whether that means big stuff that is used infrequently as well, and CONSIDER whether your entire applications pool of stored procedures is a reasonable amount of memory to consume as pinned.

Your mileage MOST CERTAINLY WILL VARY, and there is a special place in hell if it exists for anyone who suggests I suggested this as a best practice.

Within reason and in some very notable high throughput environments that I am legally prohibited from documenting this can be a HUUGE savings and a more stable environment. It *may* even be useful dovetailed with a planned monthly or quarterly pair of restarts if some memory heavy stored procedures only run on month, quarter, and year end.

Packing these things all in pinned so that everything that fluctuates and is not pinned tends to be contiguous can be a big advantage this method may implement. INCLUDE THE SYS AND SYSTEM STUFF that you use.

It is NOT that Oracle won't sort of do all this for you. But this remains something a good operational DBA can do that would be very difficult for Oracle (or anyone) to automate well for the general case. There are definitely cases where the effort is worthwhile.

Good luck,

mwf

PS: This ancient advice is NOT obsolete due to Oracle's memory tradeoff automation. In fact the balance point of it being worthwhile is tipped toward more often being worthwhile as this thread underscores.

-----Original Message-----
From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org] On Behalf Of Jonathan Lewis
Sent: Monday, February 04, 2019 5:16 AM
To: Martin Klier - Performing Databases GmbH Cc: Oracle-L Freelists
Subject: Re: ASMM - resizing triggers/thresholds

Two possibilities -

1. as you suggest, a test is not made immediately after the flush shared pool, with the result that the shared pool has started to fill by the time the test is run and the reparsing that occurs persuades Oracle that a shrink would be a bad idea.

or maybe

b) when you flush the shared_pool items that are pinned are not flushed, and if you have a lot of pinned items scattered across the shared pool then you may have some items in every granule so that granules can't be re-allocated. (But I think in this case you might see "deferred" shrink/grow reported in v$sga_resize_ops / v$memory_resize_ops. If the whole granule can't be re-allocated you may still see buffer cache activity in the shared_pool memory with v$sgastat showing memory with the name 'KGH: NO ACCESS'.

If you can flush the shared pool can't you follow up with a manual resize of the db_cache_size ? As a test, if nothing more ?

Regards
Jonathan Lewis

---

From: Martin Klier - Performing Databases GmbH <martin.klier_at_performing-db.com>
Sent: 04 February 2019 08:59:08
To: Jonathan Lewis
Cc: Oracle-L Freelists
Subject: Re: ASMM - resizing triggers/thresholds

Hi Jonathan,

thank you for the ideas!

Maybe this mechanism is not triggered by flushing the shared pool. Do you think that's possible?

Regards

--
Martin Klier // Performing Databases GmbH Managing Partner // Senior DB
Consultant Oracle ACE Director

martin.klier_at_performing-db.com // https://www.performing-databases.com



________________________________


Von: "jonathan" <jonathan_at_jlcomp.demon.co.uk>
An: "Oracle-L Freelists" <oracle-l_at_freelists.org>, "Martin Klier"
<martin.klier_at_performing-db.com>
Gesendet: Samstag, 2. Februar 2019 15:03:43
Betreff: Re: ASMM - resizing triggers/thresholds Martin,

If you look for the hidden parameters like "%memory_broker%" that gives you
some clues.
12.2.0.1

NAME                                          SES_VAL
--------------------------------------------- ------------------------------


_automemory_broker_interval                   3

_memory_broker_log_stat_entries               5

_memory_broker_marginal_utility_bc            12

_memory_broker_marginal_utility_sp            7

_memory_broker_shrink_heaps                   15

_memory_broker_shrink_java_heaps              900

_memory_broker_shrink_streams_pool            900

_memory_broker_shrink_timeout                 60000000


_memory_broker_stat_interval                  30





I believe MMAN handles the memory management - so you could enable wait
state tracing on that process in a sandbox.

I think there's a calculation every 3 seconds to determine whether or not
memory allocations should change, although strangely the parameters above
suggest both 3 seconds and 30 seconds . The former apparently is for AMM
while the latter is for ASMM.

The marginal_utility suggests to me that the calculation involves the shared
pool / buffer cache advice regarding time saved by changing memory size (a
variant of the details shown in the AWR), with a requirement (perhaps) of a
minimum improvement (12%, 7%) in performance before a change will be
triggered.

There is a parameter _memory_management_tracing, default 0, that might
enable tracing of the feature(s).

Regards
Jonathan Lewis







________________________________________


From: oracle-l-bounce_at_freelists.org <oracle-l-bounce_at_freelists.org> on
behalf of Martin Klier - Performing Databases GmbH
<martin.klier_at_performing-db.com>
Sent: 31 January 2019 10:00:35
To: Oracle-L Freelists
Subject: ASMM - resizing triggers/thresholds

Hi listers,

how and when are memory areas like DB cache and shared pool being resized,
when the instance runs in Automatic Shared Memory Management?

In my case, I have a system that has had some shared pool eating (child
cursor issue, blowing up the library cache), which was resolved. Now we have
a de-facto undersized buffer cache and low library cache consumption, but I
can't see any resizing taking place. I have SGA max size = SGA Target and no
minimum values for any component here.

Platform: Linux, Oracle EE 12.2.0.1

It would not be a big problem to restart the instance at some point, but
this may be different for other cases. So my interest is more educational.
:) In short: Where are the thresholds, what does trigger the resizing, how
often is it scanned, when does it happen?

Thanks in advance!
--
Martin Klier // Performing Databases GmbH Managing Partner // Senior DB
Consultant Oracle ACE Director

martin.klier_at_performing-db.com // https://www.performing-databases.com

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


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

Received on Mon Feb 04 2019 - 16:05:22 CET