2 Apr
2013
2 Apr
'13
6:15 p.m.
On 3/31/2013 7:33 PM, Fred Cisin wrote:
Built, used for four years, and now DISMANTLED?
What is wrong with their fundamental architecture?
<snip>
Nothing really. People learn. Four years is a long time at the high
end of computer science.
I worked on the first BlueGene (BlueGene/L) to make the list. Not too
long after that machine fell off the list RoadRunner made the top of the
list. (I had a lot of friends who worked on RoadRunner. Both were IBM
machines.)
BlueGene/L was quite a bit different than the NEC machine that held the
top slot before it. There were 64K "compute nodes" and something on the
order of 1K "I/O" nodes. Both variants used an embedded style PowerPC
system on a chip, but they were wired a little differently. Besides the
massive number of processors there was a lot of thought put into the
communications between processors ("tree" and "torus") to make it
suitable for parallel workloads that still required some coordination.
RoadRunner was a much different beast. It was essentially IBM QS22
Blades (Cell processor with enhanced double precision support compared
to the variant used in the Sony PS3) and another blade (AMD x86 I
think). Two QS22s and the other blade were packaged together in what we
called a "tri-blade", which served as the basic building block. The x86
blade handled I/O and management and left the CPU intensive work to the
QS22s.
It's really easy to throw rocks a something you don't understand. Some
light reading on Wikipedia would have shown you how much different the
different #1 machines are from each other.
Wait until we see the "exascale" style machines ...