Hi all,
I'll soon have to administer a 48 core production server, with Debian stable on it. This computer will be essentially used for scientific calculations by several users, and I have to install and mantain a fair queueing system to handle users' calculation requests, as I've seen on several computing grids. I'd like users to login and submit calculations to a queue, that should handle the sharing of cpu time and memory between users.
Which (free) software package should I use? Any advice on system and queues setup?
I have good experience with linux in general, but this is the first time I have to perform a similar task.

Thank you

You might take a look at the TORQUE Resource Manager based on OpenPBS. It's available on many distributions as binary package (package name torque and related ones).

Thank you, looking around I also found the slum+maui/moab combination, but both seem too complex and hard to manage. Isn't there something simpler? After all, I have to handle a single machine

If you can "partition" the workloads you could look at cgroups and just let the normal Linux scheduler (CFS these days) handle things.
Say users in group "A" can have 12 CPUs and a quarter of the memory - group "B" similar, but different CPUs say.
Group "X" can have everything.

Users are assigned to a group as they login - all work they generate will be assigned to the same group automagically unless you move it somewhere else. All based on PID. I've used it for years to isolate my testing workloads from my session I monitor the tests with - it needs its own small environment.

I would suggest to use SGE, which is now under developement as an open source fork at Univa after SUN was bought by Oracle and converted to proprietary software. The criticism you raise is fairly understandable. Reading 700 pages or more of a manual for installing a queuing system is nothing anyone likes to do. The queuing systems being it Torque, SGE, LoadLeveler, ... were designed for distributing jobs to remote nodes - something you won't need, as you have the special case of the qmaster being also the final exechost. Running it local is a special case. I do this even on all workstations of the students, as they can start with small computations local on their dual-core machines and using the "real" cluster later on won't change much for them.

The advantage of SGE is, that it can be installed during evaluation also under a normal user account (you can just run jobs only from this account then). If you are happy with it, you can install it again under the root account to make it available for all users. As Univa is just in the transition process, there are no compiled executables for now of the latest open source version. But you can use one of the former free open source versions from SGE6.2u4

Mailing list: gridengine.org
Wiki: Wiki