[Beowulf] large array to run
Robert G. Brown
rgb at phy.duke.edu
Thu Dec 13 15:34:49 PST 2007
On Fri, 14 Dec 2007, Chris Samuel wrote:
> On Fri, 14 Dec 2007, Robert G. Brown wrote:
>> Which I say only semi-sarcastically. They sound like they're
>> coming, they're coming. Who knows, maybe they're here and I'm just
>> out of touch.
> Dunno how close to not being vapourware these are (they may already be
> available from the last part of the article):
Ya, reminiscent of Duke's old trapeze project -- build a cluster that
just acts like a virtual or otherwise extension of memory in a NUMA
model of one sort or another.
The question there is what is the bandwidth, what sort of latency, how
does one manage e.g. locking. Should one REALLY parallelize the
project, and split the array up on a real cluster (and use the cluster
nodes to manage parallelized portions of e.g. matrix operations on the
array) or try to use a single processor with a huge virtual memory
attached via a network or use a multiprocessor with a huge virtual
memory attached via a network. Oh my aching head. Maybe we should just
try shrinking the size of the matrix by an order of magnitude in each
dimension and live in the 4 GB or so THAT would take...;-)
>> The company sells a Violin 1010 unit that holds up to 504GB of DRAM
>> in a 2U box. Fill a rack, and you're looking at 10TB of DRAM.
> of course it's not cheap..
>> Those of you who want to try Violin's gear now can get a
>> 120GB ÿÿstarter kitÿÿ for $50,000.
Let me run right down to the bank...;-)
Maybe if my novels ever sell eight million copies and I can buy boxes
like this just to play with them in my garage...;-)
Robert G. Brown
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Lulu Bookstore: http://stores.lulu.com/store.php?fAcctID=877977
More information about the Beowulf