Slashdot Mirror
Linux And Los Lobos Supercomputer
DocRea writes: "Using Linux, IBM and the University of New Mexico will connect 256 two-processor IBM Intel-based servers with high-speed Myrinet cards to
create a 512-processor machine capable of 375 billion calculations
per second. The computer, called Los Lobos, will primarily be used
for scientific purposes, but will be adapted by IBM to provide the
"cluster" approach to running software for business tasks and e-
commerce. "
The UNM sports teams are The Lobos, which is Los Lobos in Spanish. New Mexico has a large Hispanic population so they were just trying to fit in.
:)
But the band was the first thing I thought of too and I live here.
The first couple of posts after I hit 25 today, that box wasn't there, so I've got a couple of unexplained 2's floating around out there, but Slashdot's been a little flakey overall today (the GUADEC Reports story and the paying bills online story didn't have the i-opener story between them, and then later they did, internal server error messages an hour later, all that fun stuff). It's there now and I'm a-usin' it. : )
I see even classic Slashdot is now pretty much unusable on dial up anymore.
It's amusing to note that Macintosh touted it's G4 multi-processor workstations as violating export restrictions by being considered a 'super-computer'. But it just goes to show how out of date US export law is. Besides, what's to prevent someone from buying a 250 port switch from Cisco, 250 workstations from IBM, and then shipping them seperately and downloading the Beowulf software seperately?
Export restrictions aren't very restrictive, taken in that light. National security most certainly isn't being enhanced any...
You would've found the answer to your question had you bothered to actually read the front page:
Figuring out why this particular system isn't on the list is left as an exercise to the reader.
AMD's processers can do multiprocessing (or could, I vaguely recall a rumor they dropped this with Athlon). The problem is that all the MP motherboards use Intel chipsets which support Intel's style of MP. AMD (and Cyrix et al?) don't have licenses to the Intel MP technology, and came up with their own -- which requires a different mobo chipset, which nobody makes because the boards would be incompatible with Intel CPUs.
With AMD's recent growth and giving Intel a run for its money on releasing faster CPUs, we may see that change.
Oh, and a Beowulf is clustered, so doesn't care what manufacturer makes the CPUs, it can even mix them. It's tightly coupled SMP machines that care.
-- Alastair
If I remember correctly, somebody took 50-75 AIX RS/6ks and clustered them a year or two ago, with pretty nice results... Of course, you could always just get an RS/6000 S80 or a Sun e10000... or go all the way to a mainframe (I'd still love to have one, though a Cray-1 with the couch would be even better...)
"It's tough to be bilingual when you get hit in the head."
The easiest way would probably be to post about Natalie's new 8-minute petrified grits, and taunt the moderators repeatedly - that way, you could get back under 25, and not have to worry about that silly 'No Score +1 Bonus' button.
<anti-karma bait>I'm far enough over that I'd have to do some serious ninja pancake runs, Trolling for Scooby Doo, or M$ advocacy to get back under 25 (though I'm certainly nowhere near the levels of some of the famous posters here)... so I'll post this at +1 and take my chances</anti-karma bait>
"It's tough to be bilingual when you get hit in the head."
Yeah they are... unless there's actually two of these at UNM... which would be nice, but little selfish.
I can see posting "IBM to announce (newfangled technology) whis year" then a month or three later "IBM released (newfangled technology)". But two days of the same announcement is a little tough...
"It's tough to be bilingual when you get hit in the head."
But the original articles are on Wired and cnet, but it's about the same system... :)
Geeky modern art T-shirts
Wouldn't it be great if this really was a whole new story about a whole new Netfinity cluster at UNM? Then they could Beowulf the two clusters together into one of the top ten supercomputers.
Oh well, just thought I'd mention thought before this thread gets wiped.
Work for Change & GET PAID!
You can cluster AMD right now! One processor per node, but then again the Intel solutions don't ramp above two/node.
Ah yes, the practical difference between a Beowulf and a supercomputer! Dig in your pocket and pull out some coins. Nickels are commodity clusters, pennies are SGI boxes and dimes are Cray. They are a collection of the same thing, namely coins/supercomputers. But what differentiates a cluster from a Cray? The intended purpose of the components. All the bits of a T3E were designed for it. In a cluster, the parts were designed for (and in most cases are) low cost workstations, hack rigged in software to run as one monster machine.
.sig: Now legally binding!
Stop trolling. Linux isn't touching solaris in the data center market. It doesn't touch it in the scalability department, and the x86 hardware to run it isn't there even if it could.
If Linux is just killing Sun then why can't they stop making money hand over fist ? The fact of the matter is that Linux scalability plain sucks, and the answer has been "we're working on that" for at least 2 years. It _is_ getting better, but it is simply not up to par with OSes designed explicitly to handle the massive multi-cpu systems made by their vendors.
Finally, "Beowulf != Scalable". Last time i looked into it, beowulf was originally very much like an MPP and had made some inroads to providing functionality of single-image computing. The applications of linux's "clustering" are still quite narrow, and chances are most people that say they want beowulf clusters don't have any clue what they'd do with them.
Scalability for most people is something different. "Hey, i've got this webserver and RDBMS that are both thread-aware and thus can benefit from SMP machines..I'll just add more CPUs and RAM to this giant box and they'll automagically get better". Linux is an extremely poor choice for that scenario. The kernel locking is still much too coarse, too many sections are still non-reentrant, and the SMP hardware linux supports isn't particuarly scalable anyhow. Linux is _so_ poor a choice for true SMP computing that that was the area that made those amusing Mindcraft comparisions between Linux and NT realistic (i.e. NOT falsified). If you bother to look, the comparisons were using Multiple CPU boxes with Multiple Network cards. It shouldn't surprise anyone that NT did much better here, as at that time (and perhaps even still today) the linux IP stack was not re-entrant. The test took advantage of the fact that NT handled multiple device instances and SMP extremely well and Linux handled it extremely poorly. NT was designed for SMP from the ground up, with linux its been a progressive hack.
People who disagree are more then welcome to come up with facts---NOT trolls, to support their arguments. Postings which make unsubstantiated claims and predictions are basically useless.
My opinions are my own, and do not necessarily represent those of my employer.
Linux is the true industry leader in regards to scalability and security
Linux is neither. UNIX like systems are arguably a poor choice for a secure operating system since they are so damn intent on providing service and flexibility. However, even amongst UNIXes Linux is no where _close_ to being the security leader. Try OpenBSD. Any mention of "security" that doesn't also include "openBSD is pretty much the most secure UNIX flavor in wide use" is at best unenlightened.
as far as scalability, see my earlier post. Linux does _not_ lead in scalability because of its poor SMP supprt and the poor scalability of the SMP hardware it can run on.
Finally, IRIX machines can and do stay up for long periods of time, and there are frankly a hell of alot more mission-critical multi-CPU irix machines than there are _total_ multi-cpu linux boxes.
My opinions are my own, and do not necessarily represent those of my employer.
When is AMD going to get off there arses and make their processors friendly to the whole multiprocessor shindig? Translation, when do we get Athlon in a beowolf or SMP? If AMD is serious about wanting to compete with Intel than they need to start working on proof of concept projects like this. Second question is what is going to be the practical difference between a supercomputer and a beowolf cluster. They are both starting to be made more and more like each other. More to the point, how do you define a supercomputer? This is very important for things like export controls.
A midrange machine with up to 30 processorss h.990929.1.html
http://www.sun.com/servers/midrange/e6500/
This is what it does when clustered.
http://www.sun.com/smi/Press/sunflash/9909/sunfla
A machine with up to 64 processorsf lash.20000114.1.html
http://www.sun.com/servers/highend/10000/
What is it good for?
http://www.sun.com/smi/Press/sunflash/2000-01/sun
SUN and clusters? http://www.sun.com/clusters/
And remeber... If you keep FUDin' around the SUN will never shine on you... ;-F
"At the end of the journey, all men think that their youth was Arcadia..." -Goethe
- Where's my karma?
- On the parking lot son!
$HOME is where the
-- silver_p
I suppose now someone's going to ask if it runs Linux... At least it's already clustered, so that throws that question out...
I don't like fish. Reverse the fish to e-mail.
Take a look at www.scali.com They provide high performance for beowulf clusters based on linux. They have really some impressive stuff running on their test clusters. It seems that they are getting close to linear scalability of informix. Add a node and the performance increases ~linear.
Why name a supercomputer after a second rate band? I mean, yes, the soundtrack to La Bamba and Desperado, but everything else they've ever done sucks.
:-)
As long as the computer plays that cool guitar intro from Desperado when it boots, I guess it's okay.
---
- Give a man a fire and he's warm for a day, but set him on fire and he's warm for the rest of his life.
It's been done. Check this link.
The right to offend is far more important than the right not to be offended. (Rowan Atkinson)
actually.. this is very much a reality. check out The Stone Soup Cluster . It uses all kinds of machines from 486's to Pentiums all in one cluster.
...and the geek shall inherit the earth...
www.linux-skunkworks.com
there already is such a thing, although it's not as closely integrated as a beowolf. it's a process-migration kernel enhancement that allows several machines (whether workstations or clusters) to share cpu loads.
check it out at mosix.org. it is especially useful for compiling large projects (i.e. xfree86).
jon
-- http://www.cerastes.org
The athlon is SMP capable, there's just not a cheap commerical memory controller that implements it yet. It's the same bus, same SMP structure as the Alphas, and they have done it, but AMD isn't producing chipsets (you'll have to ask VIA).
"It's tough to be bilingual when you get hit in the head."