Обсуждение: pb with big volumes

On 8/10/23 16:36, Marc Millas wrote:
> Hi,
>
> I have a 15 TB db on postgres 14 (soon 15).
> shared buffers is 32 GB.

Does the system have 128GB AM?

>
> It's a db with max 15 users and often less, and currently 1 or 2.
> the biggest table have 133 partitions of 150M to 200M+ rows each.
> lots of request access explicitly one of those.
>
> When I, alone, run a query "reading" 15M buffers, it takes 20 minutes 
> (+-5minutes). inside the query there are  2 group by on a 200M rows 
> partition, with all the rows in each group by.
> When a colleague run the same kind of request (not the same request, but 
> something reading roughly the same volume ) , on a different set of data,  
>  his request is completed in less than half an hour.
> If we run our requests simultaneously... my request take hours. around 3 
> hours.
>
> I am making a supposition that its some kind of "pumping" effect in the 
> cache.

What is work_mem set to?

When were the tables last vacuumed and analyzed?

Good index support?  (But that might not matter if every row in the table is 
in the GROUP BY.)

Clustering the tables, and using BRIN indices might help.

>
> I cannot have access to the underlying OS. I can, for sure, do some copy 
> xx from program 'some command',  but its a container with very limited 
> possibilities, not even 'ps'.
> So I would like to monitor from inside the db (so without iostat and the 
> same)  the volumes of read that postgres do to the OS.
> I did activate track_io_timing, but the volumes I get in the explain 
> analyze buffer are roughly the same alone or not alone. (the 15M buffers 
> told )
> to my understanding, the volumes that are shown in pg_stat_database are 
> the useful ones ie. even if the db as to read it from disk more than once. 
> true ? or false ?
>
> So.. either my supposition is not correct, and I will read with a lot of 
> interest other ideas
> either its correct and I would like to know how to monitor this (in the 
> current context, installing a dedicated extension is not impossible, but 
> is a very boring process)
>
> Thanks for your help :-)
>
> regards,
>
> PS: I know that providing the complete data model and the exact requests 
> can be considered mandatory, but when I change the request I get the very 
> same behaviour...



-- 
Born in Arizona, moved to Babylonia.

I think your concern is that 20 min + 30 min does not equal 3 hours.

It might be natural to think the contention would, at max, be 50 min x 2 (1 hr 40 min).

So what's going on?

It seems disk I/O  is a primary suspect since you hint for an iostat replacement inside of Postgres.

If it is due to disk I/O the resolution will be to add RAID 0 SSDs at best.

Consider looking at io stats on the container's persistent volumes.

What is the pipe connecting the database server to the disks?  If it's NAS, well that would explain it.

HTH,

Adam

On Thu, Aug 10, 2023 at 2:37 PM Marc Millas <marc.millas@mokadb.com> wrote:

Hi,

I have a 15 TB db on postgres 14 (soon 15).

shared buffers is 32 GB.

It's a db with max 15 users and often less, and currently 1 or 2.

the biggest table have 133 partitions of 150M to 200M+ rows each.

lots of request access explicitly one of those.

When I, alone, run a query "reading" 15M buffers, it takes 20 minutes (+-5minutes). inside the query there are  2 group by on a 200M rows partition, with all the rows in each group by.

When a colleague run the same kind of request (not the same request, but something reading roughly the same volume ) , on a different set of data,   his request is completed in less than half an hour.

If we run our requests simultaneously... my request take hours. around 3 hours.

I am making a supposition that its some kind of "pumping" effect in the cache. 

I cannot have access to the underlying OS. I can, for sure, do some copy xx from program 'some command',  but its a container with very limited possibilities, not even 'ps'.

So I would like to monitor from inside the db (so without iostat and the same)  the volumes of read that postgres do to the OS.

I did activate track_io_timing, but the volumes I get in the explain analyze buffer are roughly the same alone or not alone. (the 15M buffers told )

to my understanding, the volumes that are shown in pg_stat_database are the useful ones ie. even if the db as to read it from disk more than once. true ? or false ?

So.. either my supposition is not correct, and I will read with a lot of interest other ideas

either its correct and I would like to know how to monitor this (in the current context, installing a dedicated extension is not impossible, but is a very boring process)

Thanks for your help :-)

regards,

PS: I know that providing the complete data model and the exact requests can be considered mandatory, but when I change the request I get the very same behaviour...

Marc MILLAS

Senior Architect

+33607850334

www.mokadb.com

On Fri, 11 Aug 2023 at 13:54, Ron <ronljohnsonjr@gmail.com> wrote:
> Wouldn't IO contention make for additive timings instead of exponential?

No, not necessarily. Imagine one query running that's doing a
parameterised nested loop join resulting in the index on the inner
side being descended several, say, million times.  Let's say there's
*just* enough RAM/shared buffers so that the index pages, once the
index is scanned the first time, all the required pages are cached
which results in no I/O on subsequent index scans.  Now, imagine
another similar query but with another index, let's say this index
also *just* fits in cache.  Now, when these two queries run
concurrently, they each evict buffers the other one uses.  Of course,
the shared buffers code is written in such a way as to try and evict
lesser used buffers first, but if they're all used about the same
amount, then this can stuff occur.  The slowdown isn't linear.

I've no idea if this is happening for the reported case. I'm just
saying that it can happen. The OP should really post the results of:
SET track_io_timing = ON; EXPLAIN (ANALYZE, BUFFERS) for both queries
running independently then again when they run concurrently.

David
David

On 8/10/23 23:40, David Rowley wrote:
> On Fri, 11 Aug 2023 at 13:54, Ron <ronljohnsonjr@gmail.com> wrote:
>> Wouldn't IO contention make for additive timings instead of exponential?
> No, not necessarily. Imagine one query running that's doing a
> parameterised nested loop join resulting in the index on the inner
> side being descended several, say, million times.  Let's say there's
> *just* enough RAM/shared buffers so that the index pages, once the
> index is scanned the first time, all the required pages are cached
> which results in no I/O on subsequent index scans.  Now, imagine
> another similar query but with another index, let's say this index
> also *just* fits in cache.  Now, when these two queries run
> concurrently, they each evict buffers the other one uses.  Of course,
> the shared buffers code is written in such a way as to try and evict
> lesser used buffers first, but if they're all used about the same
> amount, then this can stuff occur.  The slowdown isn't linear.

But that's cache thrashing (which was OP's concern), not IO contention.


-- 
Born in Arizona, moved to Babylonia.

On Mon, 14 Aug 2023 at 11:14, Marc Millas <marc.millas@mokadb.com> wrote:
> that's exactly  my question.
> does the analyze buffers data, generated when track_io_timing is on, keep track of multiple reloads of the same data
whileexecuting one operation ?
 

Yes, the timing for reads will include the time it took to fetch any
buffer that wasn't found in shared buffers.  Some of those may come
quickly from the kernel's page cache, some might come from disk. If
some other running query has evicted a buffer that the query has
previously used, then that's going to cause another pread, which will
be timed by track_io_timing and added to the count of buffers read in
the "BUFFERS" EXPLAIN output.

So, the BUFFERs EXPLAIN option showing similar amounts of reads
between the query running without the concurrent query and with the
concurrent query does not necessarily mean more buffers had to be
loaded from disk, just that fewer were found in shared buffers.  The
amount of time doing I/O as shown by track_io_timing is going to be
more interesting as that's really the only indication from within
PostgreSQL that you have to get an idea of if the buffers are coming
from the kernel's cache or from disk.  You'll probably want to
calculate the average time it took to get 1 buffer for each query to
make sense of that.

David