[Beowulf] [OT] HPC and University IT - forum/mailing list?

Mark Hahn hahn at physics.mcmaster.ca
Tue Aug 15 16:45:07 PDT 2006


> I'm not 100% sure about that Mark. I care about big-A administration. I care 
> about showing departments what resources are actually available. I care about 
> what is the most efficient use of limited University resources. When I meet 
> with researchers they often say that they had no idea that there were 500+ 
> processors dedicated to research here.

I was unclear - communication is great, and "outreach" is one of the 
things I spend a fair amount of time on.  that's not what I meant by 
big-A, though - I meant people who sit behind a big desk behind a big 
secretary who decide that today is the day to solve all our problems 
by standardizing on a single desktop for the whole university.  or that 
Outlook will be the only mail client in use.  these approaches are often
justified as "putting the uni on a sold, business-like footing".


> knowing what resources we each provide our users. Does a given organization 
> focus on hardware support, software support or both?

IMO, HPC _must_ be provided by a single group, both hw and sw.


> Running one of the new Sun x4100's with both dualcore processors at 100% uses 
> <270 watts (as determined by kill-a-watt. That is Big-A because it means that 
> we can be more efficient in our use of AC and power. It is small-a for the

but the big-A approach would ignore the fact that dual-core is very bad
for some applications.  what I'm criticising is actually when administrators
don't realize they don't know anything.  that's what's so dangerous about 
letting administrators negotiate with vendors.

power-efficiency of clusters is pretty cut-and-dry: run at as low a clock
as you can, as low a voltage, put as much onchip as possible, minimize 
infrastructure you don't need.  really smart big-A would hold brainstorming
sessions on, for instance, how to use machinerooms to heat domestic water,
or whether clocks can be modulated down a bit during peak $/KW times, or 
whether omitting per-node disks can save money.

heck, in general, how do you maximize science/research throughput of a
cluster while minimizing costs?  I can see why big-A would be interested in
this, but I don't see what wisdom they could provide.  some kind of incentive?

> Another that is both is what submission systems we are using and Why?
>
> Same questions, that affect both administration and Administration.

huh?  what value does big-A have to add here?  the correct queueing system is
the one that is cheap, low-maintenance, efficient, easy to use, etc. 
those are things that users and sysadmins know, not behind-desk-sitters...



More information about the Beowulf mailing list