Bogdan Costescu wrote:
>On Mon, 22 Mar 2004, Karl Forner wrote:
>
>
>
>>I deal with a lots of fast mpi jobs (sometimes just a few seconds) from
>>many differents users.
>>
>>
>
>Are these short MPI jobs much faster than short non-parallel jobs ?
>
>
yes
and moreover, it's difficult to predict if the job will be short or not,
it depends on the input, the parameters and so on...
>
>
>>These jobs are can be very fast so that it is no-sense starting a new
>>lam session at each time (on my cluster a lamboot usually takes around 5
>>seconds).
>>Moreover I do not want to create as many accounts as I have different
>>users, so I only use one lam-session.
>>
>>
>
>And how do you protect users' data from one another ?
>
>
I have a grid-like system (like globus, pbs, grid engine ...) on top of
it that takes care of that. users never call mpirun explictly.
actually, they do not have to know that their jobs run on a linux
cluster, it's transparent
>If you have small jobs, it usually makes sense to boot LAM only once
>per user and leave the daemons running until that user finishes all
>jobs.
>
>
in this case there is no concept of termination. when they want to run a
job they run it.
|
