On Tue, 16 Nov 2004, Leif Nixon wrote:
> Ian Stokes-Rees <[log in to unmask]> writes:
>
> > Why would I, as the cluster user (or grid user) care that two processors
> > "happen" to be in the same 1U box with the same power supply, motherboard,
> > and local harddrive(s). I suppose I can answer that myself: in the case
> > of gang scheduling and truly parallel jobs which use multiple processors.
>
> Two words: Memory bandwidth!
... but as a Grid user you don't, as your job might just as easily end up
at a resource with discrete COTS boxes with really lousy chipsets.
If "Memory bandwidth!" per se is a critical measure then there should be a
way to specify it in the JDL directly.
If (as I assume) you mean "how fast will a memory-access-heavy job
actually run" for resource broking purposes, then there might be
implications for how sites are benchmarked and their capabilities
are described (a Grid development issue), but I agree with Ian that for
independent simple jobs it shouldn't matter to users how a site actually
provides the promised resources, as long as it does.
On the basis that just because I haven't heard of something doesn't mean
it doesn't exist, are there any security issues particular to HT? I'm
thinking of user jobs trying to snoop on each other, which maybe has
received less attention than a user job trying to snoop on the kernel,
which might be an issue for sites running biomed + other jobs.
Henry
--
Dr. Henry Nebrensky [log in to unmask]
http://www.brunel.ac.uk/~eesrjjn
"The opossum is a very sophisticated animal.
It doesn't even get up until 5 or 6 p.m."
|