[distcc] Multiple people running distcc at once, using one DISTCC_DIR

[Daniel Kegel]

Jake McGuire wrote:
> File locking over NFS is a complex operation, much more so than on a 
> local disk, and distcc can do a whole lot of it.  Depending on the 
> network and the load on the cluster, I'd not be surprised if the file 
> locking took more time than actually distributing the .i file and 
> getting back the .o file.

Right.  nfs locking has been broken from the beginning (i.e. for about 20
years now); one should not rely on it.

> Using a hardware loadbalancer seems like the most robust solution, but 
> I've got no idea how much those cost.

No need for a hardware loadbalancer.  You could write a software
one that would do fine.  Essentially, using a software loadbalancer,
you could totally get rid of the need for the NFS locks.
You'd move part of the distribution logic from distcc into the
loadbalancer.  I've been thinking about this for some time,
and I like the idea.  However, it's probably overkill for
Victor's immediate needs.

Victor, could you please at least try my suggestion of
throwing out the slow machines from the cluster,
treating all remaining machines as equivalent speed,
using a nonshared local distcc directory,
and see how well that does?  I know it's a lot simpler
than what you'd been planning, but that's a good thing,
if it can achieve nearly the same performance.
- Dan