Re: Built-in connection pooling

On 01/19/2018 06:03 PM, Claudio Freire wrote:
>
>
> On Fri, Jan 19, 2018 at 1:53 PM, Konstantin Knizhnik

> <k.knizhnik@postgrespro.ru <mailto:k.knizhnik@postgrespro.ru>> wrote:
>
>
>
>     On 19.01.2018 19:28, Pavel Stehule wrote:
>>
>>
>>             When I've been thinking about adding a built-in connection
>>             pool, my
>>             rough plan was mostly "bgworker doing something like
>>             pgbouncer" (that
>>             is, listening on a separate port and proxying everything
>>             to regular
>>             backends). Obviously, that has pros and cons, and probably
>>             would not
>>             work serve the threading use case well.
>>
>>
>>         And we will get the same problem as with pgbouncer: one
>>         process will not be able to handle all connections...
>>         Certainly it is possible to start several such scheduling
>>         bgworkers... But in any case it is more efficient to multiplex
>>         session in backend themselves.
>>
>>
>>     pgbouncer hold all time client connect. When we implement the
>>     listeners, then all work can be done by worker processes not by
>>     listeners.
>>
>
>     Sorry, I do not understand your point.
>     In my case pgbench establish connection to the pgbouncer only  once
>     at the beginning of the test.
>     And pgbouncer spends all time in context switches (CPU usage is 100%
>     and it is mostly in kernel space: top of profile are kernel functions).
>     The same picture will be if instead of pgbouncer you will do such
>     scheduling in one bgworker.
>     For the modern systems are not able to perform more than several
>     hundreds of connection switches per second.
>     So with single multiplexing thread or process you can not get speed
>     more than 100k, while at powerful NUMA system it is possible to
>     achieve millions of TPS.
>     It is illustrated by the results I have sent in the previous mail:
>     by spawning 10 instances of pgbouncer I was able to receive 7 times
>     bigger speed.
>
>
> I'm sure pgbouncer can be improved. I've seen async code handle millions
> of packets per second (zmq), pgbouncer shouldn't be radically different.
>

The trouble is pgbouncer is not handling individual packets. It needs to
do additional processing to assemble the messages, understand the state
of the connection (e.g. to do transaction pooling) etc. Or handle SSL.