[Beowulf] Roadrunner shutdown
Prentice Bisbal
prentice.bisbal at rutgers.edu
Mon Apr 1 08:10:06 PDT 2013
Interesting. I know Roadrunner is still near the top of the Top500 list,
so I did a quick check of it's standings. Here's how Roadrunner compares
to Cielo, it's replacement, based on the the November 2012 Top500 and
Green500 lists:
Roadrunner:
Top500 ranking: #22
Top500 Rmax: 1042 TFLOP/s
Green500 Ranking: #148
Green500 MFLOPS/W: 444.35
Total Power: 2,345 KW
Cielo:
Top500 ranking: #18
Top500 Rmax: 1053 TFLOP/s
Green500 Ranking: #195
Green500 MFLOPS/W: 278.98 KW
Total Power: 3980
Clearly there's something more to this story. Cielo is only 1% faster
than Roadrunner, and produces 38.2 % LESS FLOPS/W. In most cases, the
energy efficiency of RoadRunner would trump the 1% performance
improvement of Cielo. I'm speculating the decommissioning of RoadRunner
is driven more by people and politics than any technical issues. I think
any of these reasons could be valid:
1. The contract for building/running RoadRunner mandates its
decommissioning at this time.
2. Support agreement with IBM has ended, and it's either too costly to
renew, or IBM isn't interested in renewing it.
3. The limited knowledgebase created by this unique system made it too
difficult to either maintain or optimize codes for it (user base is too
small), casuing it to be underutilized. (This could be argued to be a
technical issue or a people issue).
4. The space occupied by RoadRunner is needed for an even bigger system
(Exascale, anyone?).
5. No more spare parts availability.
6. Secretly being moved to South Korea to piss off North Korea even more
than the B-2 bombers did.
Any other ideas, or does anyone know the real reason driving this? I'm
sure HPCWire or some other news source will provide a more detailed
explanation soon.
Prentice
On 04/01/2013 06:45 AM, John Hearns wrote:
> I now we have all seen this through different sources.
> However interesting as this is a main stream news item, and it is
> pretty balanced.
>
> http://www.bbc.co.uk/news/technology-21993132
>
More information about the Beowulf
mailing list