[Beowulf] Marenostrum Layout

Mark Hahn hahn at physics.mcmaster.ca
Thu Apr 14 08:18:56 PDT 2005

> To see the Marenostrum layout in detail, you can see it at the
> IBM Linux on POWER site here:
> http://www-1.ibm.com/servers/eserver/linux/power/marenostrum/hpc.html
> Click on Begin in the flash demo at the top to view the room layout.

egads, IBM has terminal constipation of the documentation.  it's impossible 
to find anythign about the dissipation of a blade.  I'm guessing it's 
about 150-200W per dual.  168p/rack.  13-17 KW/rack.

it's interesting to contrast HP's blades.  we looked at their bl35p
(dual-opteron) - it has admirable density (192p/rack), and opterons 
are both reasonably fast and reasonably cool.  but if you actually
follow the blade mentality (density matters), you quickly wind up with 
racks dissipating 20+ KW (and that's not even counting the possibility
of dual-core, dual-socket blades!)  depopulating racks to make them 
easier to cool seems rather silly.

consider that for 20KW, you basically need 7 tiles per rack to 
provide cold airflow.  you might be able to get that down to 3 or so
with high-perf tiles, over-rack coolers, etc.

AFAIKT, with current dissipations, blades are only viable for machinerooms
which need a late-life kicker - say, put a single rack of blades in a room
filled with low-density stuff (and crank up the airflow some).  rows of 
fully-populated blade racks seem to be a real dissipation issue.

remember that a 1U dual-socket, dual-core server is 4p/u - the same density
as marenostrum.  of course, it may not be desirable to divide interconnect 
among 4p...

now, if we could only get vendors to agree on a standard, flexible heatpipe
connection scheme ;)

More information about the Beowulf mailing list