[pgpool-hackers: 3246] Re: select(2) storm

[Jesper Pedersen]

Hi,

On 2/18/19 10:57 AM, Jesper Pedersen wrote:
> However, if we extended pgpool with a "state machine" which keeps track of
> 
> 1) In which "direction" to read from, and hence write to
> 2) Known states where we need to change "direction"
> 
> we should see a sizable speedup. If proxy.c is compiled with 
> -DSTATE_MACHINE I see a 20%-50% performance improvement. I tested with 
> the various pgbench scenarios; -M simple, -M extended, ...
> 
> The "state machine" just keep track of the last client message, and last 
> server message seen in order to "pair" them - f.ex. P -> 1, B -> 2 and 
> so on.
> 

Here is an updated version.

This version can deal with server streams such as notifications.

The way it is handled is by using O_NONBLOCK on the sockets, and 
switching "direction" upon EAGAIN or EWOULDBLOCK. This is faster than 
using select(2) plus the extra logic it would take to keep track of this.

I have successfully used this version against various work loads, and is 
even seeing an increased performance, up to 95%.

I'm thinking about if we need select(2) at all...

There is a changelog in the file.

Thanks !

Best regards,
  Jesper
-------------- next part --------------
A non-text attachment was scrubbed...
Name: proxy.c
Type: text/x-csrc
Size: 11614 bytes
Desc: not available
URL: <http://www.sraoss.jp/pipermail/pgpool-hackers/attachments/20190219/66465759/attachment.bin>