While I violently agree with your analysis in the general case, and because
I have the same view of the outcome either way, I'll take that bet. This is
a new system, so the existing plan is the only plan, not a new plan.
Indeed, continuing your medical analogy, suggestions to alleviate the
current systemic bottleneck should have the immediate impact of keeping the
patient alive at a higher functioning state while a cure is sought for the
underlying problem. So even if the "allopathic" course makes the system
appear healthy, the "homeopathic" followup should be pursued.
Slapping in the parallelism and reduction in frequency of allocation fixes
in this case are likely cheap and effective to at least improve the
throughput. Again, since the system is new, and has apparently only been
subjected to high current load with the "go-live" process is seems quite
likely to me that a fair number of commonly constructed queries may be
spawning large hash plans and/or sorts, so it is less likely the real fix
will be a quick fix of a single query. (Of course you're right that one
missing index or a poorly-conceived index could generate a lot of problems.
It is also entirely possible we're seeing a ramp up in data volume of a new
system where the nearly empty stats generated all manner of fts/hash plans
that will become good plans as soon as the stats are updated, and the former
tiny in memory hashes and sorts flashed by but now dump to disk.)
So take the two aspirin (for the pain) and see if you can't find a way to
stop hitting your thumb with the hammer. Now if it wasn't the rough
equivalent of two aspirin (as well as being responsive to the OP's actual
question), but rather was an expensive thing to relieve the symptoms while
he searches for the real cure, I'd vote for some interim load control while
the underlying cause(s) is/are cured. But spreading out temp access and
reducing the allocation frequency should be cheap and relatively effective.
... just don't stop there.
:) mwf (Where am I buying the 1st or 2nd round of drinks where you're buying
the 2nd or 1st round of drinks?)
-----Original Message-----
From: oracle-l-bounce@xxxxxxxxxxxxx [mailto:oracle-l-bounce@xxxxxxxxxxxxx]
On Behalf Of Tim Gorman
Sent: Monday, April 28, 2008 11:31 AM
To: goran00@xxxxxxxxx
Cc: hkchital@xxxxxxxxxxxxxx; oracle-l@xxxxxxxxxxxxx
Subject: Re: 'buffer busy waits' on Header Block (#2) of Tempfile
I agree with Stephane's response in principal as well as specifics.
Responses on this thread have fallen into one of two camps:
1. address the symptom as described
2. find out more about what is going on, because the
observed effect may be based on an anomaly in the application
In the medical arena, course #1 can be described as "allopathic" while
course #2 can be described as "homeopathic". On dictionary.com,
"homeopathy" is defined as "the method of treating disease by drugs,
given in minute doses, that would produce in a healthy person symptoms
similar to those of the disease", while "allopathy" is described as
"the method of treating disease by the use of agents that produce
effects different from those of the disease treated". In essence,
homeopathy is the more considered response (i.e. "why is symptom XYZ
happening"), while allopathy is the knee-jerk response (i.e. "treat
symptom XYZ with response ABC"). Each has their appropriate
application in medicine, but in database and application tuning, I
would argue that the homeopathic course is more effective in the
larger number of cases. Find out what is going on, why so much TEMP
space is being demanded, because the reason could be something simple
like a missing or poorly formed index.
In fact, I will gladly bet a round of drinks that the problem is a new
execution plan recently adopted by the CBO or a missing index, or a
poorly-conceived index, or a combination of all three, and has nothing
to do with the configuration or implementation of the temporary
tablespaces, nor does it have anything to do with PGA configuration.
...I'm glad to be wrong as well as right, because either outcome still
results in a round of drinks...
Any takers?
Quoting goran bogdanovic <goran00@xxxxxxxxx>:
Hi,
Since your contention is on header block, increasing extent size from 256k
to 1mb sound reasonably to me.
One of the reasons why processes checks the segment header is to extend
the
HWM.
Have you found which queries are consuming much of the temp space? Can
they
be optimized?
regards,
goran
On Mon, Apr 28, 2008 at 2:36 PM, Hemant K Chitale
<hkchital@xxxxxxxxxxxxxx>
wrote:
We'd rolled out a new system today and user connections rapidly
went up to 900 sessions. OK, we are supposed to be able to handle that.
However, by the afternoon, I saw 'buffer busy waits' on Block#2 of the
first tempfile of the Temporary Tablespace (using a custom Temporary
tablespace instead of 'TEMP', although the default 'TEMP' exists).
By evening, we had 300 sessions waiting on 'buffer busy waits' on the
same
block
(querying V$SESSION_WAIT for P1, P2).
Although the tablespace has 3 tempfiles, the first file has 350 sessions
against it and the other two have less than 100 sesssions put together
(querying V$TEMPSEG_USAGE for SEGFILE#).
These are a mix of SORT and HASH extents.
How can I address this ?
a. Increase PGA_AGGREGATE_TARGET (1GB for 400 concurrent users + 10-15
batch jobs)
and/or switch to WORKAREA_SIZE_POLICY='MANUAL' with SORT_AREA_SIZE and
HASH_AREA_SIZE.
b. Add more tempfiles to the tablespace and/or rebuild the tablespace
with
larger
Extent Sizes (1MB instead of the current value of 256K)
Any other suggestions ?
Hemant K Chitale
http://hemantoracledba.blogspot.com
"A 'No' uttered from the deepest conviction is better than a 'Yes' merely
uttered to please, or worse, to avoid trouble."
Mohandas Gandhi Quotes :
http://www.brainyquote.com/quotes/authors/m/mohandas_gandhi.html
--
//www.freelists.org/webpage/oracle-l
--
//www.freelists.org/webpage/oracle-l
