From: | Charles Nadeau <charles(dot)nadeau(at)gmail(dot)com> |
---|---|
To: | Claudio Freire <klaussfreire(at)gmail(dot)com> |
Cc: | Igor Neyman <ineyman(at)perceptron(dot)com>, Jeff Janes <jeff(dot)janes(at)gmail(dot)com>, "pgsql-performance(at)postgresql(dot)org" <pgsql-performance(at)postgresql(dot)org> |
Subject: | Re: Very poor read performance, query independent |
Date: | 2017-07-18 09:20:35 |
Message-ID: | CADFyZw4m8e5dO-FXqW0Xcf1NP_P=7Vn7qCfNrb__puzgTiPxEA@mail.gmail.com |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-performance |
Claudio,
Find attached the iostat measured while redoing the query above
(iostat1.txt). sda holds my temp directory (noop i/o scheduler), sdb the
swap partition (cfq i/o scheduler) only and sdc (5 disks RAID0, noop i/o
scheduler) holds the data. I didn't pay attention to the load caused by 12
parallel scans as I thought the RAID card would be smart enough to
re-arrange the read requests optimally regardless of the load. At one
moment during the query, there is a write storm to the swap drive (a bit
like this case:
https://www.postgresql.org/message-id/AANLkTi%3Diw4fC2RgTxhw0aGpyXANhOT%3DXBnjLU1_v6PdA%40mail.gmail.com)
I can hardly explain it as there is plenty of memory on this server. The
execution time of the query was 4801.1s (about 1h20min).
I reduced max_parallel_workers_per_gather to 2 and max_parallel_workers to
3, restarted postgresql then ran the query again while running iostat again
(iostat2.txt): The query ran much faster, 1992.8s (about 33min) instead of
4801.1s (about 1h20min) and the swap storm is gone! You were right about
the max_parallel_workers_per_gather!!
For the last test, I changed the scheduler on sdc to deadline (iostat3.txt)
keeping max_parallel_workers_per_gather=2 and max_parallel_workers=3 then
restarted postgresql. The execution time is almost the same: 1938.7s vs
1992.8s for the noop scheduler.
Thanks a lot for the suggestion, I'll keep my number of worker low to make
sure I maximize my array usage.
Charles
On Mon, Jul 17, 2017 at 10:56 PM, Claudio Freire <klaussfreire(at)gmail(dot)com>
wrote:
> On Fri, Jul 14, 2017 at 12:34 PM, Charles Nadeau
> <charles(dot)nadeau(at)gmail(dot)com> wrote:
> > Workers Planned: 12
> > Workers Launched: 12
> > Buffers: shared hit=728798037
> read=82974833
> > -> Hash Semi Join
> > (cost=30059688.07..47951761.31 rows=220376 width=20) (actual
> > time=1268845.181..2007864.725 rows=7057357 loops=13)
> > Hash Cond: (flows.dstaddr =
> > flows_1.dstaddr)
> > Buffers: shared hit=728795193
> > read=82974833
> > -> Nested Loop
> > (cost=0.03..17891246.86 rows=220376 width=20) (actual
> time=0.207..723790.283
> > rows=37910370 loops=13)
> > Buffers: shared
> hit=590692229
> > read=14991777
> > -> Parallel Seq Scan on
> flows
> > (cost=0.00..16018049.14 rows=55094048 width=20) (actual
> > time=0.152..566179.117 rows=45371630 loops=13)
> > Buffers: shared
> > hit=860990 read=14991777
> > -> Index Only Scan using
> > mynetworks_ipaddr_idx on mynetworks (cost=0.03..0.03 rows=1 width=8)
> > (actual time=0.002..0.002 rows=1 loops=589831190)
> > Index Cond: (ipaddr
> >>=
> > (flows.srcaddr)::ip4r)
> > Heap Fetches: 0
> > Buffers: shared
> > hit=589831203
>
> 12 workers on a parallel sequential scan on a RAID-10 volume of
> rotating disks may not be a good idea.
>
> Have you measured average request size and average wait times with iostat?
>
> Run "iostat -x -m -d 60" while running the query and copy a few
> relevant lines (or attach the whole thing). I suspect 12 parallel
> sequential scans are degrading your array's performance to random I/O
> performance, and that explains the 10MB/s very well (a rotating disk
> will give you about 3-4MB/s at random I/O, and you've got 2 mirrors on
> that array).
>
> You could try setting the max_parallel_workers_per_gather to 2, which
> should be the optimum allocation for your I/O layout.
>
> You might also want to test switching to the deadline scheduler. While
> the controller may get more aggregate thoughput rearranging your I/O
> requests, high I/O latency will severly reduce postgres' ability to
> saturate the I/O system itself, and deadlines tends to minimize
> latency. I've had good results in the past using deadline, but take
> this suggestion with a grain of salt, YMMV.
>
--
Charles Nadeau Ph.D.
http://charlesnadeau.blogspot.com/
Attachment | Content-Type | Size |
---|---|---|
iostat3.txt | text/plain | 15.9 KB |
iostat2.txt | text/plain | 16.3 KB |
iostat1.txt | text/plain | 38.8 KB |
From | Date | Subject | |
---|---|---|---|
Next Message | Claudio Freire | 2017-07-18 16:01:29 | Re: Very poor read performance, query independent |
Previous Message | Claudio Freire | 2017-07-17 20:56:23 | Re: Very poor read performance, query independent |