I don't really want to open the debate again but I must say that this is
one of the few occasions that I really, truly and utterly disagree with
Steve. ;-)
<rant>
First off, by any reasonable definition of "physical" rather than
"logical" CPU a core is a physical CPU. If I went out and bought some
worker nodes with 4 CPU slots and installed single core CPUs in all of
them as far a Grid jobs and scheduling goes that would be almost
indistinguishable from a MoBo with two Dual core CPUs or one with one
quad core CPU. And if they are that sensitive then they should be
worried about the motherboard chipset, the speed of the FSB and any
number of other hardware choices that could have equally miniscule
effects on jobs.
Then by using physical/logical names for CPU/core distinction it leaves
us with nothing to use for HyperThreading, which is what the logical CPU
was originally thought up for and is of course coming back with Nehalem,
or overloading Nodes with more jobs than CPUs.
And all that's before we get into how well or badly clusters and
subclusters are supported by YAIM etc. The whole setup strikes me as
something that looked OK in theory but takes no account of what is
actually done.
The LRMSs are some of the most mature and sophisticated pieces of code
we have (significantly predating the Grid) but we cripple them by giving
then no information about the jobs passing through.
Yes, subclusters should be useful for gross differences like
architectures, operating systems but using the for minor differences
between worker nodes just adds complexity to the sites (extra queues, if
nothing else) for, as far as I can see, no benefit.
</rant>
Yours, getting it of his chest,
Chris.
> -----Original Message-----
> From: Testbed Support for GridPP member institutes [mailto:TB-
> [log in to unmask]] On Behalf Of Ewan MacMahon
> Sent: 20 November 2009 14:39
> To: [log in to unmask]
> Subject: Re: Cores, CE_PHYSCPU, CE_SMPSIZE, CE_LOGCPU
>
> > -----Original Message-----
> > From: Testbed Support for GridPP member institutes [mailto:TB-
> >
> > Why the long face? That's what sub-clusters are for.
> >
> The long face is because of the subsequent discussion about the
> lack of support in YAIM, the lack of support in the WMS, and
> the fact that if Stephen and Steve don't seem to have entirely
> congruent understandings of something then deploying it is
> likely to be messy.
>
> And yes, that does mean I was depressed about this conversation
> before it happened; did anyone just turn something on that might
> be warping the space-time continuum or otherwise altering the
> fabric of reality?
>
> Ewan
--
Scanned by iCritical.
|