[GE users] Interactive / PE
andreas.schwierskott at oracle.com
Thu Jul 22 11:13:03 BST 2010
[ The following text is in the "utf-8" character set. ]
[ Your display is set for the "ISO-8859-10" character set. ]
[ Some characters may be displayed incorrectly. ]
> I have a few questions regarding functionnalities of SGE. I have assumed a
> lot while reading the doc, and I guess it needs to be corrected now.
> I only have a month left to finish my configuration. It doesn't need to be
> optimized, but it has to be easy to use and to mod in the future, and
> should require minimal maintenance.
> 1/ I read somewhere ( I can't find it anywhere now ...) that interactive
> sessions were much better left alone on a node. Several interactive
> sessions are ok on the same node, but mixing interactive + seq. or
> parallel would diminish perfomance. Is this true and should I take this
> into account for a recent cluster?
if the sum of all CPU-bound processes equals the number of cores on a host
virtually every additional CPU activity can heavily impact the performance
of a parallel job. This goes that far that some sites need to reduce the
sge_execd PDC interval (by default once per second, in total contributing
less than 1% to the CPU usage, see PDC_INTERVAL in sge_conf(5)) to get as
little unwanted load as possible on a host.
Other sites go even further and analyze the justification of every single
daemon running on a host to ensure best possible parallel job performance.
You find more in this topic by searching for "Operating System Jitter" / "OS
Jitter" in you favorite search engine.
> 2/ If a user choses to work 'out of core' (i.e. on the disk), can another
> heavy (but a bit less so it works incore) be launched on the same nodes ?
> How does the user specify he wants to work out of or in core from the
> 3/ I am configuring parallel environments :
> - Is it the users responsability to use MPI / OpenMP, irregarding the
> - selected parallel environment ? My parallel queues are set up 1*16
> - nodes, 2*8 and 4*4. Do I only need 3 PE (one for each 'type' of queue) ,
> - in which only the slots number will change and where slots = available
> - cores? Then the users submits his job with OpenMP or OpenMPI (and maybe
> - sometimes with OpenMP in OpenMPI)
> - Is it bad to have an odd (uneven) number of slots for parallel
> - application, as they work by pairs ? I want to leave one core free to
> - run the execd on each server in case it takes too much %CPU.
> - As for the allocation rule, I think $round_robin is the fair solution,
> - even if slots could be 'lost', it's still better than $fill_up (for MPI
> - ofc, OpenMP would use $fill_up).
> - Also, as my parallel queues are superordinated to all other queues which
> - they share nodes with, max_reservation is not used. I think
> - control_slaves has to be set to true for a tight integration.
> Is this set up correct ?
> I found much doc for OpenMPI, but much less for OpenMP. Is there a reason for this ?
> Thanks for having read,
> Le nouveau Hotmail est presque arrivé, ne le manquez pas !
> To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].
To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].
More information about the gridengine-users