[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Building a cluster with debian?

On Wed, 3 Dec 2008 14:18:37 +0200
Micha Feigin <michf@post.tau.ac.il> wrote:

> On Wed, 3 Dec 2008 12:40:45 +0200
> Johann Spies <jspies@sun.ac.za> wrote:
> > On Tue, Dec 02, 2008 at 04:53:45PM +1100, Alex Samad wrote:
> > 
> > > I believe there is software out there that will do that, I read about
> > > some recently, but can't find reference to it now :(
> > 
> > Have a look at http://gridengine.sunsource.net/
> > 
> Thanks, looks like a glorified pbs mostly but may be of some help. I was
> hoping to somehow abuse numa and processor queues to make the cluster look as
> a single numa computer but that seems to require hardware support (multi
> processor numa system and not several systems combined into one numa).
> The solutions I found so far:
> 1. Hardware - Buy a very expensive multi opetron system (not an option at
> this time)
>2. ssi (single system image) cluster, seems to be what I want,
> it's just not clear to me at the moment if there is anything that can
> properly migrate threads http://en.wikipedia.org/wiki/Single-system_image

Looks like current options are
LinuxPMI ( continues that dropped openmosix project)
Kerrighed (web page says that they are planning to allow thread migration in november)
Mosix 2 (commercial)

Looks like all of them support comunication through pipes, doesn't seem like
any of them support thread migration yet, would love to hear otherwise.

> 3. Batch/Grid systems
> Your reference seems to fall into the third category.
> The problem with this is that you either need to use mpi or cluster openmp
> (which I haven't managed to get working on non-opetron systems) or to run
> several unconnected jobs in parallel.
> I prefered number 2 but as I said, it doesn't seem to handle multi threaded
> and shared memory programs, most of them not even as a whole (i.e once you
> have threads the whole job will not migrate).
> I want to get more threads for matlab. The solution is matlab
> distributed computing (which costs about 4000$, not cheap also), and agian,
> mostly allows managing several jobs in parallel on a cluster but not a
> massively parallel job
> I will read a bit more into these as the machines are supposed to arrive
> tomorrow finally (4 quad core machines). Will try to give a followup if anyone
> is interested.
> > Regards
> > Johann
> > 
> > 

Reply to: