Re: profiling connection overhead

On Mon, Dec 6, 2010 at 12:38 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> One possible way to do make an improvement in this area would be to
>> move the responsibility for accepting connections out of the
>> postmaster.  Instead, you'd have a group of children that would all
>> call accept() on the socket, and the OS would arbitrarily pick one to
>> receive each new incoming connection.  The postmaster would just be
>> responsible for making sure that there were enough children hanging
>> around.  You could in fact make this change without doing anything
>> else, in which case it wouldn't save any work but would possibly
>> reduce connection latency a bit since more of the work could be done
>> before the connection actually arrived.
>
> This seems like potentially a good idea independent of anything else,
> just to reduce connection latency: fork() (not to mention exec() on
> Windows) now happens before not after receipt of the connection request.
> However, I see a couple of stumbling blocks:
>
> 1. Does accept() work that way everywhere (Windows, I'm looking at you)

Not sure.  It might be useful to look at what Apache does, but I don't
have time to do that ATM.

> 2. What do you do when max_connections is exceeded, and you don't have
> anybody at all listening on the socket?  Right now we are at least able
> to send back an error message explaining the problem.

Sending back an error message explaining the problem seems like a
non-negotiable requirement.  I'm not quite sure how to dance around
this.  Perhaps if max_connections is exhausted, the postmaster itself
joins the accept() queue and launches a dead-end backend for each new
connection.  Or perhaps we reserve one extra backend slot for a
probably-dead-end backend that will just sit there and mail rejection
notices; except that if it sees that a regular backend slot has opened
up it grabs it and turns itself into a regular backend.

> Another issue that would require some thought is what algorithm the
> postmaster uses for deciding to spawn new children.  But that doesn't
> sound like a potential showstopper.

The obvious algorithm would be to try to keep N spare workers around.
Any time the number of unconnected backends drops below N the
postmaster starts spawning new ones until it gets back up to N.  I
think the trick may not be the algorithm so much as finding a way to
make the signaling sufficiently robust and lightweight.  For example,
I bet having each child that gets a new connection signal() the
postmaster is a bad plan.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company