[distcc] Re: RFC: new protocol draft

Martin Pool mbp at samba.org
Wed May 7 02:40:17 GMT 2003


On  6 May 2003, Ben Elliston <bje at wasabisystems.com> wrote:
> Martin Pool <mbp at sourcefrog.net> writes:
> 
> > I had in mind that there would be a long-lived client process started
> > implicitly when you first connect to a host.  It will hold open
> > SSH connections and pass them out as needed.  Connections are passed
> > between the long lived process and distcc invocations using
> > file-descriptor passing over a Unix domain socket.
> 
> > Sound good?
> 
> Sure.  Mike Stump has also been discussing a similar idea on the GCC
> list, whereby a GCC "server" hangs around with parse trees of common
> header files in core.  A whacky, but interesting idea.

Isn't this more or less what the Apple precompiled header thing does,
except that it mmaps them to disk rather than keeping them in a process?

> I was also wondering if you had thought about TTCP.

Transactional TCP, to avoid the three way handshake?

I don't think the handshake is a big deal, as we will almost always be
on a low-latency link, and the handshake is small compared to the
amount of data to be transmitted.  (So we're unlike the T/TCP
best-case of small conversations with a remote server.)

In any case my impression was that T/TCP is dead.  (I think Rusty said
that.)  There doesn't seem to be an implementation in the Linux 2.4
stack.

e.g. 

  http://www.xent.com/FoRK-archive/feb99/0255.html

If it was in the kernel then I might turn it on to see if it helps,
but since it's not there I don't think it's worth worrying about.

As for HTTP-NG, I'm going towards reusing a stream rather than
modifying TCP.

-- 
Martin 


More information about the distcc mailing list