Was this a paid for advertisement?
I don't see any balance in the article.
Dell has claimed it is simplifying supercomputing by replacing complex monolithic, proprietary architecture Cray-like machines with modular ones, using racks of industry-standard components. In fact, in one way it's helping to complicate supercomputing, because writing parallelised code is so hard. But the massed ranks of x86 …
"This assumes that needed data sets could be sent to the Dutch HPC centre somehow." .... Or its Head be a Source Sharer ....... IntelAIgent Farmer/Star Pro Phormer. Ideally a Being with HyperVision at Least Matching ITs Uncanny Abilities with Virtual Facilities.
And it would be madness not to expect CyberSpace to be an Environment of Great Game Interest to TUDelft and Noordwijk ESA...... for Advanced IntelAIgent Virtual Defence.
x86 based super computing by putting up hundreds of servers in a cluster is not really super computing is it? When DoD, NASA, NOAA, universities, etc., have an HPC requirement, they go to IBM, Cray or SGA, not Dell. There's a lot more to super computing than massive parallelism.
"Just 1% of electronic failure". This means waht? 1% of servers had a failure? or 1% of electronic components gad a failure?
As for buying new kit and having more processor power for less money... well, it is expected that 5 years old kit is slow.. and that's why most people consider it legacy.
As for dell providing more for less, this is very old news. But I can also buy similary priced HP Proliant kit.. or even Sun x86 (a bit slower and more expensive).. as for reliability.. well, if you want reliability buy Sun/Fujitsu T2 machines.. more quality, half the components, faster and more expensive.
Does the pope shit in the woods? (or something)
We're replacing our Sun E25k machines with Sun 5440 for fraction of the cost and does the job faster and cheaper (no dynamic reconfiguraition, but not such a loss as it never worked properly anyway)
"Sun E25k machines with Sun 5440 for fraction of the cost"
You sound like one of my junior SEs. For awhile it will. But a 5440 will never hold up the the pounding an E25K is designed to take. And Niagera chips' single-treaded performance is abysmal. But US4+ isn't much better. You're right about DR. Sun never has got that right. Their virtualization is pretty weak, too.
A common misconception is that supercomputing is all about cycles.
That commodity boxes can outperform dedicated systems (like CRAY)
Really, the issue is more about the type of data sets and how tightly coupled the data elements are relative to each other.
If you have a loosely coupled problem (like rendering frames in an animated film), then your interprocessor / memory communication needs are relatively light and you can get away with stacks of commodity processors merrily churning away Mickey Mouse (no pun intended) images. If you are trying to model turbulent fluid flow , nothing beats a good, dedicated vector processor.
Having owned and run a CRAY J90 and constantly wagered to "beat the LINUX HPC cluster", even the old workhorse (which is a pretty dated unit ) is up to the challenge. No wonder the black helicopter folks choose the big iron constantly.
Programming a CRAY-level system can be "challenging" and the operational carbon footprint is enough to bring tears to one's eyes, ( not to say a shock when the electricity bill comes due).
For more information (and debate) try this site:
Paris, because she knows all about multitasking (hey, had to thow that in...)
This article was such rubbish it's not worth a rant. But seriously... Computers get faster every six months. Not a stretch a replacement box is cheaper and faster... Duh. It bothers me when vendors brag someone switched from ancient hw to newer hw by them and it performed better. I should hope so!
Mike's spot on. Even going from an older uber Sun box to one of their newer ones give performance increases and costs less. And the Sun Niagra2 boxes are cheaper to run then a larger cluster of x86 boxes...
"Most issues in science today are computational in nature,"
Nope. Most of the issue is data management--how do I represent scientific data and manage it over its lifecycle, and how do I track the relationship between observed data and derived data, etc.
No amount of cycles and storage can solve that problem, which is why it remains unsolved decade after decade.
I have to agree with previous comments. This article appears to be a case of a writer not very familiar with the HPC arena getting a bit carried away by Dell smoke. Old enthusiasms for commodity clusters touted as if they are new ideas; comparisons between new and old machines cited as if they are the least bit relevant; not a single thought as to why so many organizations shell out millions when apparently a 60K pile of Dell boxes would do just fine... definitely has the scent of Texas marketers trying to fast talk a room full of suits.
>>You sound like one of my junior SEs.
hehe.... flamebait? (25 years in IT, worked on everything from SS4's to E25s and the newer T5440s, superdomes, ICL mainframes, IBM 590s, Sun, Oracle & Cisco certified)
>> For awhile it will. But a 5440 will never hold up the the pounding an E25K is designed to take.
What does this mean? will a T5440 wear badly? this is the kind of marketing junk line that managers swallow. In reality it all *Depends on your workload* but they aren't that far apart in performance.
I'd argue that a fully spec'd T5440 will outperform a fully spec'd E25 using the same SAN for any real-world business processes, for a fraction of the orginal cost, more threads, faster memory architecture (OK only 512Gb rather than 1Tb but hey!), E25s are crippled by the non uniform memory architecture and scale badly over 5 processor boards, the T5440 doesn't suffer from this, most companies with big E25s either chop them up into smaller domains or have to optimise for them (we shouldn't have to do this).
>>Their virtualization is pretty weak, too.
"pretty weak" nice vague statement again with no substance, the Sun zone/container/domain model is one of the most complete (except perhaps for IBM LPars, but don't get me started, 0.1 of a proc? no fine processor affinity?). We've saved hundreds of thousands with virtulising the Sun kit, it runs faster, cooler, cheaper and so much more flexibly (not that weak).
And a s a final note.... M9000 (faster than a P595, I think I just creamed myself)
"This, Claman claimed, was half the compute power of the human brain."
Okay, stop right there. We haven't a clue what the compute power of the human brain is. We haven't a clue how it works. We don't even know what it does. (What is intelligence? Seen any definitions recently that lend themselve to an objective test of whether intelligence is present or not, and if so how much? Thought not.)
Therefore, anyone who compares this or that computer to the human brain is mentally freewheeling and anything else they might have said should be treated with the utmost caution.