[distcc] distcc 0.13 and "Default task limit"

[Scott Lystig Fritchie]

Greetings!  I've been using distcc for a couple of months now and have
really found it to be quite valuable.  I started playing with distcc
0.13 today and noticed something a bit peculiar with 0.13's "distcc"
behavior when picking a remote machine that didn't happen with
previous versions.  Then again, my situation is a bit peculiar.

"distcc" seemed to do a good job (with its local locking scheme
mucking around in /tmp) of choosing a DISTCC_HOST to send a job to.
However, that doesn't work if I have several machines running "make".
So, I wanted to insert a TCP proxy in between my "make" build machines
and my back-end "distccd" machines.

I was dissatisfied with the TCP proxies that I could find in Open
Source Land.  Most of them were written for application-level firewall
or HTTP proxying use.  Those that load-balanced did simple round-robin
or some variation.  My back-end machines vary quite a bit in terms of
CPU speed: I didn't want the proxy choosing a slow back-end when a
faster one was available...

... so I wrote my own TCP proxy.  It chooses the fastest back-end
machine that happens to be idle.  It nicely deals with multiple-CPU
back-end boxes.  It's written in Erlang, if that interests you or
bothers you.  If anyone is interested in its code, I can see if my
employer would mind letting me release it under free/free'ish license.
{shrug}

Anyhow, I'm leading up to my newly-discovered problem.  I just started
using 0.13 today.  I did the following:

    % setenv DISTCC_HOSTS "proxy proxy proxy proxy proxy proxy proxy proxy proxy proxy proxy proxy proxy proxy proxy proxy proxy"
    % make -j17 blahblah

And I discovered that the proxy only had 4 jobs active at one time.
The "make" box is an AMD XP1600+, and even though the distcc + cpp
overhead is pretty low, this machine can't actually get 17 going very
often.  But 8-12 is no problem.  Seeing only 4 was a big surprise.

I found something in 0.13's src/hosts.c's dcc_parse_hosts() function:

        /* Default task limit */
        curr->n_slots = 4;

This, and later usage of n_slots elsewhere, appears to limit the
number of jobs that any single host can perform.

I've simply fallen back to using 0.12.  I suppose I could add a few IP
aliases to the proxy box and use DISTCC_HOSTS="proxy1 proxy1 proxy1
proxy1 proxy2 proxy2 proxy2 proxy2 proxy3 ...", but I don't think I'll
go there quite yet.

I assume that this "default task limit" is a feature that isn't
implemented fully yet?

-Scott