Re: pgbench unable to scale beyond 100 concurrent connections

Hi Fabien,

Sorry for very short report.
I feel pgbench is not so complex tool.

Please see below answers to your questions.

On Wed, Jun 29, 2016 at 5:07 PM, Fabien COELHO <coelho(at)cri(dot)ensmp(dot)fr> wrote:

>
> Hello Sachin,
>
> Your report is very imprecise so it is hard to tell anything.
>
> What version of client and server are you running?

I am testing it with 9.6-beta1 binaries. For server and client it is same.
I am using pgbench on top of postgres_fdw.

> On what hardware ? (200 connections => 200 active postgres processes, how
> many processes per core are you expecting to run?

I am running in small virtual box machine. with 1GB RAM and 2 cores.
I think there should not be problem with 200 processes on 2 core machines.
I tested same number of concurrent connections on same machine with
sysbench it is working fine.

I am not sure what is difference between pgbench and sysbench, they might
be process based or thread based.
But I can say if I can create 200+ concurrent connection with sysbench , i
should be able create same with pgbench.

Thoughts?

> the recommanded value is about 2 connections per physical core...)

I think 2 connections per core is very small value . for 200 i need atleast
100 core machine , which is not good.

What precise command is started?

> How to you know it "comes down to 100 connections"?

I put watch on live connections to database.
something like : watch -n 1 'ps -ef | grep postgres | grep 192.168.56.101 |
wc -l'

NOTE: grep cxommand may change as per environment.

Also user below query to see active connection.
# select count(*) from pg_stat_activity;

> Are there error messages from pgbench or postgresql?
>
> postgresql does not give any error.

pgbench says:
client 36 aborted in state 2: ERROR: could not connect to server "server_1"
DETAIL: FATAL: sorry, too many clients already

> My random guess would be that you start too many connections with only one
> thread client side and/or on a too small hardware client or server-side for
> the expected scale, so given the load and latency some connections just
> never get to do anything?
>
> This may be reason but it should be able to maintain idle connection for
that time if never get to do anything.

> Maybe try with "-j 20" so that there are not too many connections per
> pgbench thread?
>
> I do not have such good hardware for now.
I feel pgbench should be able to perform well on small hardware.

Feel free to ask any question regarding setup.

> I am testing pgbench with more than 100 connections. also set
>> max_connection in postgresql.conf more than 100.
>>
>> Initially pgbench tries to scale nearby 150 but later it come down to 100
>> connections and stable there.
>>
>> It this limitation of pgbench? or bug? or i am doing it wrong way?
>>
>> ---
>> I tested it with max_connection = 200 in postgresql.conf
>> and pgbench witn -c 180/190/200
>>
>
> Please reply.
>>
>
> Please send precise information instead of expecting people to guess...
>
> --
> Fabien
>

--

Thanks and Regards,
Sachin Kotwal