Re: Performance issue - high Free buff wait

Thank You Jonathan.

My mistake, I didn't notice the KEEP and other pools allocation of buffer cache were hardcoded.

As per attached the AWR section from buffer pool statistics and streams activity. Buffer pool statistics does show , it's the default buffer cache which is experiencing all the "free buffer waits". Zero "free buffer waits" logged against Keep and recycle pool. Then moving down to segment by block changes , I see two of the table partitions + index partition at the top. And the "sql order by execution" does map to the sqls doing row by row insert and Update involving those tables.

Seeing no data exists under section "Memory Resize Ops" and "Memory Resize Operations Summary" for both issue and non issue period. So I hope it means no memory movement happens in/out of the buffer cache.

CPU and other resources noted under the "Streams section" of the AWR report are under ~10minutes of DB time. So I hope that may not be a contributing factor here.

With regards to total memory being ~500GB+ and 60% of this is free on this host. Currently for this database we have ~24GB allocated to SGA and ~12GB for PGA_Aggregate_target. So by what amount of additional memory should we bump the SGA_MAX_SIZE here?

I am now seeing there are ~19 tables which are assigned to the KEEP pool(which is ~8GB in size) , but the sum of the size of all those ~19 tables are coming around ~13GB. And the odd thing is one of those tables is exceeding ~12Gb itself in size. So i need to check if it's really serving some purpose or we can remove the KEEP pool and merge it to the default buffer cache only.

Regards

Pap

On Mon, Jan 18, 2021 at 2:15 AM Jonathan Lewis <jlewisoracle_at_gmail.com> wrote:

> I made two mistakes reading the details about parsing -
> a) I forgot the "parse time" in the profile was in hundredths of seconds
> b) I took the library cache mutex X at 2,200 seconds to be reasonably
> consistent with the 3600 figure for parse time, but since that was actually
> 36 seconds, and since it wasn't CURSOR mutex X I was putting 2 and 2
> together to get 5.
> So the parsing thing is probably mostly irrelevant.
>
>
> Don't be fooled by the fact I typed MB instead of GB - you have 512GB of
> memory and a huge fraction of it is "free". Why isn't more given to the
> database.
> We're also beginning to see some detail of what you're up to: your title
> is High Free Buffer Waits, but it's taken about 10 posts before you mention
> that you've got a KEEP cache and a DEFAULT cache, and that there are
> several GB of SGA which won't be available to the shared_pool or buffer
> cache because they are allocated to the Streams Pool. (Find out what
> you're application is supposed to be doing in that 15 minutes and ask
> yourself whether that might have some extra impact because of Streams;
> check the Streams section of the AWR report to see if there is any Streams
> activity going on at the time, or at all).
>
> Since you've got a KEEP cache and a DEFAULT cache, can you identify which
> cache is subject to the free buffer waits. (There's a very old article
> about the oddities that used to appear with multiple caches here:
> https://jonathanlewis.wordpress.com/2006/11/21/free-buffer-waits/ that
> might help, but there's a "Buffer Pool Statistics" section of the AWR
> report that gives you a simple summary.) Knowing which objects are
> associated with the KEEP pool may also be helpful. And the "Segments by
> ..." sections of the AWR will give you some information about which objects
> are subject to most changes in the period. Combine that with "SQL ordered
> by ..." (Executions is the obvious starting point) to get some idea of
> what's happening in the interval if you have no other way of finding out
> what the application is supposed to be doing. Bear in mind that "%
> captured" is reported for several of these reports, so it's important to
> check if you're looking at a meaningfull percentage or whether some figure
> looks harmless because it's only a small percentage of the real total.
>
>
> Another thing to check (which MAY be related to the library cache mutex X)
> is the section on memory resizing - do you have much activity of granules
> of memory being moved between the library cache the streams pool, and the
> buffer cache.
>
> Regards
> Jonathan Lewis
>
>
>
> <http://www.avg.com/email-signature?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail> Virus-free.
> www.avg.com
> <http://www.avg.com/email-signature?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail>
> <#m_-2580758597909172818_m_-2322454506161063817_DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2>
>
> On Sun, 17 Jan 2021 at 19:31, Pap <oracle.developer35_at_gmail.com> wrote:
>
>> Thanks much, Jonathan.
>>
>> With respect to parsing , i have attached the "time model
>> statistics' section (for issue period and normal period in two tabs) and it
>> shows , the parse time is <1% of the over all DB time and even the top sql
>> by parsing , the top one which is shown as ~2.7million execution does use
>> bind variable, its something as this. (SELECT MAX (col1) FROM TAB1 WHERE
>> ID1 = :B2 AND date1 = :B1;). So then, can we safely say parsing(hard
>> parsing) to not be one of the causes?
>>
>> To your point "are you "*wasting" a lot of memory (and CPU) to manage a
>> file-system. If not, how hard are the other applications / databases on
>> this machine working during this period*."
>>
>> Not heard of any issue from other databases exactly during the same point
>> in time but yes they have very less activity happen on them and i do see a
>> small spike in concurrency during that exact time on them too. But , if I
>> see the host memory utilization pattern from the last 7 days it's always
>> lying below ~40%. And the CPU utilization is max ~75% but it's not showing
>> a spike during that ~15minutes window. I am checking this on the OEM Host
>> utilization section , so not sure if it's hiding anything here. And as I
>> mentioned, the spike in AAS in the top activity section of this database is
>> clearly visible daily for that ~15minutes window in the top activity
>> section of OEM which touched to ~80(~64 being the limit).
>>
>> Also Jonathan to your point *"and your SGA size is only 24MB so you may
>> be trying to recycle your buffer cache incredibly quickly.", *the SGA is
>> actually set ~24GB , however if i see some of the major components hard
>> coded and are in GB's , they are as 1) Keep buffer cache set as ~8GB ,
>> 2)Default buffer cache - 6GB, 3)Shared pool - 4Gb and 4)streams pool- 5GB.
>> So does it mean that we really need to bump up the size of SGA so that the
>> default buffer cache can get more memory which may help us to reduce some
>> contention because the default buffer cache recycle frequency can be
>> minimized ?
>>
>> Regards
>>
>> Pap
>>
>> On Sun, Jan 17, 2021 at 9:31 PM Jonathan Lewis <jlewisoracle_at_gmail.com>
>> wrote:
>>
>>> 1) No.
>>> 2) No
>>>
>>> Going back to first principles - typical Unix systems time-slice CPUs
>>> 1/100 of a second at a time and OLTP systems do tiny amounts of work that
>>> are likely to take far less than 1/100 of a second if they're written
>>> efficiently. So a well-written OLTP system can probably support a few tens
>>> of user sessions per (real) CPUs because most of the session will spend
>>> most of their time not doing anything.
>>>
>>> If, however, you have 64 CPUs and 64 sessions which are trying to model
>>> protein folding images from a chemical formula then you've allocated too
>>> many sessions because every single one of them will want 100% of a CPU 100%
>>> of the time and you've got no CPU left for anything else. More
>>> realistically, if you have a couple of sessions which are trying to do
>>> something resource intensive like running a query over the last 2 years of
>>> sales data to produce a one page summary of customer trends then they are
>>> probably taking out one CPU per session (or N CPUs each if you run the
>>> queries parallel N). So rule 1 is "know your application"
>>>
>>>
>>> Now take a look at the two bits of AWR you first posted.
>>> 1) The average single block read time in BOTH time periods is around 1ms
>>> - which means its coming from cache or flash; so it's likely to be more CPU
>>> than I/O time.
>>>
>>> 2) The number of db block changes in the heavy period is about 60,000
>>> per SECOND, compared to about 11,000. (Halve these figures since roughly
>>> half will be for undo records going to pinned blocks) - so in the busy time
>>> period you are potentially make 30,000 blocks dirty per SECOND and will
>>> need to write them out, which is about 240MB per second, and your SGA size
>>> is only 24MB so you may be trying to recycle your buffer cache incredibly
>>> quickly.
>>>
>>> 3) As a side effect of the db block changes you are generating 12.7MB of
>>> redo per second compared to 2.8MB per second. In a production system not
>>> only does that have to be written to at least 2 files, it has to be re-read
>>> and copied to an archived log destination.
>>>
>>> Combining 2 and 3 it's not surprising that the rate of data change means
>>> you have a huge amount of conflict in (a) writing data (b) finding free
>>> buffers to read more data and (c) waitinig for other users to read the
>>> data you want to read.
>>>
>>> It's not easy to determine cause and effect after this - but the rate at
>>> which you work in that 15 minutes is simply overloading the system. So the
>>> first thing I'd do is try to find out what that 64 CPUs / 64 Cores really
>>> means. HP/UX is one of the less well-known operating systems, so any
>>> comment you hear about "unix systems" might be very misleading - but to me
>>> it looks as if you may not really have 64 CPUs. Then I'd ask about the I/O
>>> configuration - I know HP/UX is a bit unusual here, but what are the
>>> setting to avoid double caching - you have 512GB of RAM and assign 24GB to
>>> this SGA - are you "wasting" a lot of memory (and CPU) to managing a
>>> file-system. If not, how hard are the other applications / databases on
>>> this machine working during this period.
>>>
>>> The other thing that stands out as a separate anomaly is the parsing -
>>> which ends up with lots of "lilbrary cache - mutex X". The counts are high
>>> in both cases and worth looking into, but very expensive in the worst
>>> case. How many of your "parse count" turn into "parse count (hard)", and
>>> what does the time model show as the distribution of parse time. At a guess
>>> I'd say that in this interval you kick of several processes (maybe around
>>> 10) that are all trying to optimize the same statements at the same time -
>>> which leads to the "mutex X" (the rough idea is that 9 out of 10 would be
>>> waiting while the 10th does the optimising) - anifd maybe there are a lot
>>> of statments which need a lot of parsing in an environment that is already
>>> generaing a lot of CPU conflict. Note that your EXECUTE count jumps from
>>> 6,400 per second to 13,200 per second - you mentioned single-row
>>> processing, the numbers suggest that that might be the case, but also that
>>> a lot of them use literal values, and several sessions may be trying to
>>> parse and execute the same statement at the same time.
>>>
>>>
>>> Summary:
>>> Check your system and how its configured.
>>> Understand what the application is doing in this 15 minutes
>>> Then worry about getting the same job done with less work and
>>> (particularly) less conflict.
>>>
>>>
>>> Regards
>>> Jonathan Lewis
>>>
>>>
>>>
>>> On Sun, 17 Jan 2021 at 11:18, Pap <oracle.developer35_at_gmail.com> wrote:
>>>
>>>> Basically I have two doubts here,
>>>>
>>>> As it seems for 15minutes period we are pushing the System beyond its
>>>> capacity. So considering we have 64 core CPU here , should we only allow
>>>> max 64 active session at any given point from application? And then
>>>> considering we have 4 app server we need to set max session limit to 16
>>>> from each app server?
>>>>
>>>> Another doubt I had was if the dB writer response time of 100ms during
>>>> that issue period is the cause or the result of slowness? If it's the cause
>>>> can asynch IO setup for data file help improve situation here? Or by just
>>>> making row by row operations converted to batch will be enough ?
>>>>
>>>> Regards
>>>> Pap
>>>>
>>>>
>>>
>>> <http://www.avg.com/email-signature?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail> Virus-free.
>>> www.avg.com
>>> <http://www.avg.com/email-signature?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail>
>>> <#m_-2580758597909172818_m_-2322454506161063817_m_-7043823991385307789_m_3768997358934924510_DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2>
>>>
>>

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




application/vnd.openxmlformats-officedocument.spreadsheetml.sheet attachment: Buffer_pool_Stream_stats.xlsx