[GE users] Implications of Hyperthread enabled CPUs and jobs
reuti at staff.uni-marburg.de
Thu Nov 27 13:37:42 GMT 2008
[ 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. ]
Am 27.11.2008 um 10:27 schrieb Neil Baker:
> I?ve been reading up on Hyper-thread enabled CPUs and am wondering
> if it is best to disable or enable Hyper-threading in the BIOS?
I switched it always off in the past. My experience was also around
up to 50% only. Depending of you applications, you might want to try
a setup with 3 slots per node (with 2 physical CPUs). Maybe one of
them in a different queue with a nice value (i.e. priority in the
queue definition) of 19 as some kind of background-queue.
> Most of our grid machines have 2x physical 3.0 GHz Pentium 4 CPUs
> with Hyper-threading enabled. We are running openSuse10.0 on them
> (BIGSMP 2.6 kernel) so they appear to be 4x logical processors.
> If 2 jobs are assigned to run on the same machine, is the kernel or
> grid engine smart enough to run them on two physically separate CPUs,
Yes, the Linux kernel scheduler should take care of it.
> or is there a random chance that they may both run on the same
> physical CPU (i.e. the main and hyper-threaded logical CPU on the
> same physical CPU) and therefore leave the other physical CPU
> totally unused?
As the Linux kernel scheduler will change the assigment over time:
yes, it might happen and you might see a system task on the other CPU.
> I?ve read that hyper-threading only gives up to 30% more
> performance, so by accidentally running 2 jobs on the same physical
> CPU it is going to drastically slow these jobs down compared with
> running them on two separate physical CPUs.
To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].
More information about the gridengine-users