[Beowulf] I'm so glad I didn't buy one of these

Prentice Bisbal prentice.bisbal at rutgers.edu
Thu Jul 3 07:36:49 PDT 2014


On 07/02/2014 05:21 PM, "C. Bergström" wrote:
> On 07/ 3/14 04:17 AM, Jeff Johnson wrote:
>> If they want to spend a bazillion dollars to run hpl faster than anyone
>> else who am I to stop them.
>>
>> If however they want to do real science perhaps they need to architect
>> something more manageable.
>>
>> They should bust that thing up into 3 or 4 clusters.
>>
>> On 7/2/14, 2:11 PM, Douglas Eadline wrote:
>>> China's world-beating supercomputer fails to impress some potential 
>>> clients
>>>
>>> http://www.scmp.com/news/china/article/1543226/chinas-world-beating-supercomputer-fails-imp 
>>>
> replied too fast the above link gets cut
> So paste "ress-some-potential-clients" at the end or here it is again
> http://www.scmp.com/news/china/article/1543226/chinas-world-beating-supercomputer-fails-impress-some-potential-clients 
>

Ahhhh.. that's much better. Disregard my earlier e-mail about the page 
not found error. Didn't look closely enough to see the original link was 
broken.

I know I'm late to this party, and haven't read through the 20+ e-mails 
on this topic, but here's my initial thoughts:

Sadly, this is not entirely surprising. Shortly after the K Computer 
went into production, Japanese scientists made statements that the K 
computer was great for LINPACK scores, but was underperforming (relative 
to LINPACK result) on actual applications they were trying to run on the 
K. I tried searching for the article I read on this a few years ago, but 
couldn't find it. Please excuse my last of citation.

Regarding China specifically, several years ago, I attended a lecture by 
Bill Tang from Princeton University. He had just come back from China, 
where he met with their computational scientists, and was sharing what 
he learned there. According to his lecture, China acknowledged that 
despite their advances on the hardware side, China was still way behind 
the rest of the world on the software side, and it would be years before 
they could catch up, which sounds like one of the contributing factors 
to this situation.

It's hard to fault China for building this system, though. If China 
wants to become an HPC superpower, they need systems like this they can 
practice on and learn from. Yes, the 'tuition' is quite expensive, but 
learning by doing (and making mistakes along the way) is a proven 
learning technique.

What's shocking is the business model. Didn't anyone do a business 
feasibility study? I would have thought the government would be 
subsidizing it's operation to help foster the learning I mentioned in 
the previous paragraph.

Prentice





More information about the Beowulf mailing list