[distcc] On Style - the hosts file

[Jean Delvare]

Hi Daniel,

[Jean Delvare]
> I would also question the interest of such an option. Was there a
> significant improvement when compared with a fully randomized
> distribution? I would think that a more simple rule would be not to
> randomize the first host of the list and start randomizing after that.
> Wouldn't it be sufficient?

[Daniel Kegel]
> Any number of approaches would do.   I didn't want to force
> randomization down anyone's throat, so I made it an option.

I do not discuss the fact that not everyone wants host randomization, so
it has to be left as an option to the user. I was merely asking what
granularity was needed for this option. I would myself be happy with a
simple switch (randomize all slots or none), but it looks like other
people have other needs (like randomizing a part only of the list). My
question was, how sure are we that such a granularity is needed? If it
turns out not to bring any significant improvement, then maybe the
simple switch approach is sufficient - and it solves the point
altogether (although in this case I would still suggest that localhost
as the first entry of the list is considered differently, i.e. is never
randomized for the sake of configure scripts).

[Jean Delvare]
> Or we could have ~/.distcc/hosts.random for randomized hosts.

[Daniel Kegel]
> Nah, let's not have two hosts lists.

You're probably right, it might be more trouble than is worth.

What about a modifier on the host line (in the hosts config file)? After
all, there are already a lot of them with a funny syntax (for limiting
the number of slots, enabling compression etc...). As we want per slot
randomization, it would probably make sense to extend the "/" modifier
for randomization. Something like:

localhost/1
alpha/4r
beta/4r
gamma/2r

Which I think is rather explicit. All slots tagged "r" would belong to
the random pool instead of being picked in order.

Would such a solution please everyone?

The avantages over the previous approach are:
1* We keep the one-host-per-line logic, no option lines.
2* You don't have to group the randomized slots in the file.

Now, again, I wouldn't want to do this if it has no proven advantage
over the simple switch approach.

Thanks,
--
Jean Delvare