Re: Very poor read performance, query independent

Joshua,

I use noop as the scheduler because it is better to let the RAID controller
re-arrange the IO operation before they reach the disk. Read ahead is set
to 128:

charles(at)hpdl380g6:~$ cat /sys/block/sdc/queue/read_ahead_kb
128
charles(at)hpdl380g6:~$ cat /sys/block/sdc/queue/scheduler
[noop] deadline cfq

Thanks!

Charles

On Wed, Jul 12, 2017 at 1:42 AM, Joshua D. Drake <jd(at)commandprompt(dot)com>
wrote:

> On 07/11/2017 04:15 PM, Merlin Moncure wrote:
>
>> On Mon, Jul 10, 2017 at 9:03 AM, Charles Nadeau
>> <charles(dot)nadeau(at)gmail(dot)com> wrote:
>>
>>> I’m running PostgreSQL 9.6.3 on Ubuntu 16.10 (kernel 4.4.0-85-generic).
>>> Hardware is:
>>>
>>> *2x Intel Xeon E5550
>>>
>>> *72GB RAM
>>>
>>> *Hardware RAID10 (4 x 146GB SAS 10k) P410i controller with 1GB FBWC (80%
>>> read/20% write) for Postgresql data only:
>>>
>>> The problem I have is very poor read. When I benchmark my array with fio
>>> I
>>> get random reads of about 200MB/s and 1100IOPS and sequential reads of
>>> about
>>> 286MB/s and 21000IPS. But when I watch my queries using pg_activity, I
>>> get
>>> at best 4MB/s. Also using dstat I can see that iowait time is at about
>>> 25%.
>>> This problem is not query-dependent.
>>>
>>
>> Stop right there. 1100 iops * 8kb = ~8mb/sec raw which might
>> reasonably translate to 4mb/sec to the client. 200mb/sec random
>> read/sec on spinning media is simply not plausible;
>>
>
> Sure it is, if he had more than 4 disks ;) but he also isn't going to get
> 1100 IOPS from 4 10k disks. The average 10k disk is going to get around 130
> IOPS . If he only has 4 then there is no way he is getting 1100 IOPS.
>
> Using the above specs (4x146GB) the best he can reasonably hope for from
> the drives themselves is about 50MB/s add in the 1GB FWBC and that is how
> he is getting those high numbers for IOPS but that is because of caching.
>
> He may need to adjust his readahead as well as his kernel scheduler. At a
> minimum he should be able to saturate the drives without issue.
>
> JD
>
>
>
> --
> Command Prompt, Inc. || http://the.postgres.company/ || @cmdpromptinc
>
> PostgreSQL Centered full stack support, consulting and development.
> Advocate: @amplifypostgres || Learn: https://pgconf.us
> ***** Unless otherwise stated, opinions are my own. *****
>

--
Charles Nadeau Ph.D.
http://charlesnadeau.blogspot.com/