[distcc] load management

[Martin Pool]

On 27 Apr 2004, Dan Kegel <dank at kegel.com> wrote:

> If fd passing is used, the speed difference might be noticable,
> since it would completely eliminate one copy and a lot of task
> switching.

Right.  But unless there is some reason for traffic to pass through a
local daemon you can just skip it altogether...

> >>It even lets us pass credentials, so the local server could
> >>even know for sure which unix user was making the request;
> >>that could come in handy if we want to restrict status info
> >>about jobs to the user who submitted the jobs.
> >
> >Or you could just chmod the (directory containing the) socket. 
> 
> No, 'cause then I'd need a separate socket endpoint for each user,
> and I'd have to worry about whether /tmp was writable
> (it's not on one system I tried to install distcc on!).

Does gcc even work without a TMPDIR?  I doubt if distcc does.

> I'd prefer to have a single server, installed by root
> (not a problem in my environment, though I understand it'd
> be a drawback in some places).

I don't want to require it.

> Your suggestion (in a separate note) of trying out fsh is interesting,
> but fsh seems to also require the data to be sent via fshd,
> which means one extra copy and one extra task switch per block.
> Plus it's written in Python, violating the 'just c' policy of
> distcc.

No, I said "something like fsh".  fsh itself is too slow.

For the particular case of talking to distccd, we could arrange for
distccd to keep listening on the socket after a job is complete, which
avoids the need to run fshd.  We just need something on the client to
hold open the fd and pass it to distcc client processes as they need
it.  Perhaps this would be a good place for a client daemon using a
unix socket, though I'm pretty sure it should be per-user.

> I'd like to go for a world record time compiling e.g. the
> linux kernel, so I'm a bit more worried about silly little
> performance issues than is reasonable or healthy :-)

  http://lwn.net/2002/0321/a/7.5-second-compile.php3

Hard to beat. :-)  But it might be fun to try.

-- 
Martin 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
Url : http://lists.samba.org/archive/distcc/attachments/20040428/06f9daf2/attachment.bin