On Sat, Jul 27, 2002 at 10:46:22PM -0300, Carlos Carvalho wrote:
> Folks,
>
> I've just discovered bproc and it looks very interesting. However I
> have a question that is crucial to our usage.
>
> At the moment we have a small cluster of identical machines. Our users
> have FORTRAN programs that right now are not parallelized and run on a
> single machine. If I understood the docs, bproc doesn't automatically
> move jobs between nodes.
That is correct. A process can only move itself.
> How then can one distribute the load among the
> nodes? The only way I see is that the users will have to launch their
> programs via bpsh, specifying the node where they want the program to
> run. However, how can one discover the load of each node? The only way
> I see is to run
>
> % bpsh -aps 'cat /proc/loadavg|cut -f1'
>
> or something similar. Is there a better way?
>
> It'd be really nice if bproc did load balancing :-)
I try to draw a pretty clear line between what BProc's job is and what
the scheduler's job is. As I see it chosing which node to place a job
on is the scheduler's job and actually putting it there is BProc's
job. In other words, BProc doesn't make those decisions because I've
decided that scheduling is a separate problem.
- Erik
P.S. I'm trying to get a simple BProc-oriented scheduler we've written
out the door.
|
