Hi Ewan,
On 01/20/2016 03:31 PM, Ewan MacMahon wrote:
> the schema not being able to tell the difference between
> hyperthreading units and actual cores (they're both 'logical CPUs'),
People are leaning towards the idea that, when we speak of "logical
CPUs", we are not talking about CPUs at all! We're actually talking
about the number of slots that the machine has that can run jobs. The
idea is that you set the number of slots to where you get the maximum
total HS06 throughput through the node as a whole (i.e. most work when
using many instances of the benchmark.) And we call that "logical CPUs".
> separate subclusters for each generation of nodes
That's a pain, and I don't do it. I munge it all together, as per this:
https://www.gridpp.ac.uk/wiki/Publishing_tutorial
The outcome is that the total job slots, total power, and scaling for
accounting are correct (ARC Bug, ticket118930, notwithstanding) but the
cost is that information on the individual "generations of nodes" is, of
course, aggregated away. No-one has ever complained, so I assume that
information is surplus to requirements. If so, the system "works".
> the whole principle doesn't work at all for non-traditional 'cloudy'
> resources
You can say that again.
> the practical point here and now is that these figures are already wrong
Yes, and no!
Yes, the internal structure of the cluster, its nodes and its CEs is
lost in the munging, but the end result is still adequate for the
purposes listed. The figures are aggregated, and information on the
individual "generations of nodes" is lost, that's true. But it is
possible to work out properly a) the total jobs slots, b) the true
accounting and c) the true power if the Tutorial is followed
religiously. If that is all that is needed from these variables, then
the system "works".
The alternative is the CLUSTER node idea, which retains the internal
structure of the cluster. I'll go there after I've seen someone go there
first without perishing!
Cheers,
Steve
--
Steve Jones [log in to unmask]
Grid System Administrator office: 220
High Energy Physics Division tel (int): 43396
Oliver Lodge Laboratory tel (ext): +44 (0)151 794 3396
University of Liverpool http://www.liv.ac.uk/physics/hep/
|