Slashdot Mirror
World's Fastest Supercomputer to be Linux
xinit was one of the people who pointed us to the CNET story running about the possibility that a current bid by SGI for a supercomputer could be run on Linux. The supercomputer could be the fastest in the world at time of its production. SGI has confirmed the bid, saying it's being targeted for 2001, if the bid is accepted. The placement would be Los Alamos National Laboratory.
Unicos runs on the Cray family of supercomputers such as the T90 and SV1 families of vector processing supercomputers (which use custom designed vector processors) or the T3E family of massively parallel supercomputers (which use DEC Alpha processors). SGI is spinning off Cray or something effectively similar to that. SGI won't be releasing any Unicos running computers in the future.
SGI also makes massively parallel computers, which if properly configured (read lots and lots of processors) are supercomputer class machines. These machines presently run, hold on to your hats, IRIX. These machines presently use MIPS processors. One of these machines is part of the ASCI contract (Accelerated Strategic Computing Initiative) and is based on an Origin 2000 system.
Right now SGI is developing CC-NUMA computers (the same multiprocessing technology behind the Origin 2000 computers) using Intel IA-64 processors. Rather than attempting to port IRIX to an Intel processor or pretending that Windows NT will scale SGI is relying on Linux. Right now Linux can't do it, but SGI is working on improving that aspect of Linux. This is all stuff thats been posted to slashdot before. Here's a blurb to that effect.
There are several Linux clusters in the top 500, not individual computers and certainly not anything that could replace an E10K easily. While Beowulf is great for a certain class of parallelizable problems. A Big Iron database server such as the E10K is useful for it's single system image and huge memory access as well as fault tolerance and failover capablities in hardware and software. I'm not convinced that Linux will ever be used in such as system, since the advantages of open source are not as great in an environment where there are few end users, and where the end users already are spending enormous amounts of money on hardware, software, and support contracts. So the cost of an OS (really just the price of the OS support contract) is minimal in comparison to other costs. And the ability of the hardware to work tightly with the OS is a major selling point.
Linux will continue to thrive in the low-end and will migrate up to more and more powerful servers as they get cheaper and used more generally. High Availability solutions are already beginning to surface, as with TurboLinux. This will probably be the way to go for most modest sized enterprise applications.
The only way Linux will get onto a Big Iron box is for SGI, IBM or Sun to put it on there. The only good reason to do this would be to ease migration from low end solutions running on Linux. Or to appease the PHB's that demand a Linux based solution (just wait...It'll happen). Since they wouldn't abandon their current customers, they would be supporting two OSes in the same space with the same developers for quite some time. While the Linux solution would be open source, there would not be great advantages to this since the user community would be so small.
--
"L'IT c'est moi!"
You know, I have to wonder, when will you people realize that it's a pipedream?
If this system is to be ready by 2001, and is to be faster than ASCI Red (9,282 Intel Pentium Pro 200MHz with 1024k cache each, using proprietary interconnects (IP over SCSI, IIRC)) then it's not going to happen.
Even if the gov't is willing to sacrifice the reliability of the system, Linux is not ready and will not be for many years. Period. You're talking of going from 2 processors, which only works on x86 and Alpha currently, to over a thousand.
Sorry, folks, it isn't going to happen. It's a matter of rapid development, and availability. Sure, SGI could probably do it, but not by due date. Nor would it perform as necessary due to the requirement of extensive assembly-level optimization in both compilers and kernels.
ASCI Blue Pacific, IBM's entry, is one of the most powerful computers in the world. And what I'm about to say will probably send most of you into denial and/or shock.
Blue Pacific isn't customized all that much. In fact, barely customized. It's nearly the same machine you can order for your business today. Perhaps even slightly slower.
That's right - ASCI Blue Pacific CTR SP Silver and ASCI Blue Pacific, #11 and #2 respectively on the Top 500 list, are retail systems with some additional software. IBM's SP Silver in Poughkeepsie is a retail SP.
Shocking, isn't it? That someone can build a supercomputer that any business can buy. IBM holds a *lot* of the first 50. You don't see Sun till 54 with a machine that was totally custom built. Hell, look at #20! IBM SP Power3 200MHz. *200MHz!* And it smokes 480 supercomputers! That should tell you something right there. Now, SGI's talking about, more than likely, an x86 supercomputer or IA64 supercomputer, that's supposed to run Linux, have more than a thousand processors, and outperform ASCI Red? Nope, 'fraid not, folks. Maybe in 5 or 6 years, but not one and a half. Sorry. Deal with it.
-RISCy Business | Rabid unix guy, networking guru
your company here.
shelby != ford
SGI has more than one line of supercomputers. ASCI Blue Mountain is an SGI Origin2000 machine. I don't think we can expect to see Linux on Crays in the near future. (And didn't SGI just divest Cray again?)
It's true that Linux hasn't currently "mastered" 16-CPU SMPs. If Larry McVoy is to be believed, that's probably a good thing for the correctness and stability of the kernel.
CPlant is number 129 on the TOP500 list; it's the fastest Linux machine currently listed that runs Linux. It used to be below 100, but more new machines were added.
The 1000-node genetic-programming cluster mentioned recently on Slashdot, and distributed.net, are not on the list at all; to get on the TOP500 list, you need to run LINPACK fast. This (a) does not interest some people, and (b) is not well-suited to the structure of some clusters. A parallel machine that is very fast for some tasks may be very poor at others.
With regard to DES cracking: the EFF's DES cracker, which cost less than a quarter of a million dollars to build, cracks DES keys in a matter of days. Such a machine can scale linearly. The fact that distributed.net takes a month to crack a single DES key does not demonstrate that the NSA requires months to do the same.
Generally, "secure" DoD sites are not connected to the Internet, auditing or no.
"supercomputer" and "enterprise server" are very different categories. "enterprise server" means "mainframe killer" -- that is, reasonable CPU speed, fast I/O, but above all, reliability. Linux is definitely fit for supercomputing, and is being used for supercomputing all over the world. Linux is probably not quite yet fit for being an "enterprise server".
However, many supercomputers do indeed need lots of disk storage.
With regard to http://www.gapcon.org/listg.html: someone said, "You will notice there are no Linux installations in that list." Actually, they list a bunch of machines from Atipa Linux Solutions at LANL, the Avalon Beowulf at LANL, the Parnass2 Beowulf at the University of Bonn, the LoBoS Beowulf at the National Institutes of Health, the Centurion Beowulf at the University of Virginia, and possibly some others. They're a minority, but they're way cheap, and they're growing fast.
With regard to the GPL: if I hack something proprietary into Linux, I need to give source, licensed under the GPL, only to people whom I give binaries to. I am under no obligation to give source to anyone else. However, the person to whom I give it can put it up on their FTP site if they want.
Kragen Sitaker, current Beowulf FAQ maintainer
The second fastest current supercomputer (ASCI Blue) resides at LANL as well. Will they be running these concurrently, or are they scrapping their current cluster to put this one in?
The current fastest, ASCI Red, is located in Sandia National Labs. Both these systems were built by Intel I believe, and are gigantic clusters running some custom software.
The biggest Linux box/cluster/whatever is Avalon I believe, currently ranked #160, and also resides in LANL. Wasn't there one in the 50-60 range as well?
Personally I can't see this being anything but a good thing for Linux, both in terms of another selling point (Hey, it's good enough to be on the world's fastest computer!) as well as (hopefully) advancements in scalability (I can't imagine SGI implementing a massive cluster of single CPU boxen, meaning they may take a long hard look at SMP code and optomize it for whatever platform they're considering rolling out for this).
And, it has to be said:
Imagine a Beowulf of these things! Heh...
No, I'm sorry folks. That line is to be read:
to be installed in Nate Fox's garage (NFg) in suburban Los Angeles.
You'd think a major publication like C|Net would get thier facts straight ;)
...Nothing is so smiple it cant get screwed up.
-----
If Bill Gates had a nickel for every time Windows crashed...
Don't get me wrong, I love Linux. But let's be realistic here. That article contained little to no factual data. The only thing they're running on is conjecture -- SGI has expressed interest in Linux in the past, so they're assuming this hundred-million-dollar multi-teraflop machine will run it? I'll believe it when I see it.
I'll bet you, unless SGI come up with some sort of Beowulf solution instead of their time-tested Cray supercomputers, we'll be seeing yet another Unicos machine at the top of the "World's Fastest Supercomputers" list in a few years.
- A.P.
--
"One World, one Web, one Program" - Microsoft promotional ad
"Remember when the U.S. had a drug problem, and then we declared a War On Drugs, and now you can't buy drugs anymore?"
Wakko is absolutely right when it comes to the relative merits of Linux and Unicos on SMP or parallel machine - Linux wasn't designed for that job and won't do it. Not now, not in the near future. Unicos currently runs machine with up to 2048 processors. Linux hasn't really mastered 16.
However, the proposed machine is bound to be a cluster, not a single machine. Linux will do this just fine. In fact Linux would run on ASCI Red if there were drivers for the networking hardware. (They've booted individual nodes with both Linux and NT).
So there is a possibility of Linux. I guess it might be easier to use Linux on IA64 than port IRIX/Unicos.
My reservation here is that neither the Itanic nor MIPS chips offer cutting edge performance.
First of all, there's quite a bit of difference between fastest and fastest known. I can't imagine both the chinese AND the american governments from having some exceedingly classified hardware that blows the pants off the open stuff(read: governmental phallus-phlashing.)
Second, the meaning of fastest is very unclear. I'd go so far to say that any system that implements a given function in software instead of hardware is going to be orders of magnitude slower than the state of the art. Witness the EFF DES cracking machine, 3D Graphics Accelerators, even Math Coprocessors. Fitting a square peg into a round hole is actually a pretty common occurance in the computer world, but it takes a relatively tortoise-like rate compared to what can be pulled off with raw gates.
That's why XISC--Extensible Instruction Set Computing--is probably the upcoming processor paradigm. Programmers need the ability to redefine round holes into square ones, so the square pegs fit right in.
Yours Truly,
Dan Kaminsky
DoxPara Research
http://www.doxpara.com
It depends on the dimension you're scaling in. Parallel procesing power is the easiest to scale in.
/netname > /ipnetaddr > /ipnetbits > /ipnetdef > < /netdef >
Mainframes aren't that much more powerful than desktops in processing power, but they are much more powerful in terms of I/O bandwidth and storage capacity.
The scalability issue that most people are talking about is scalability on an enterprise network in number of users and diversity of missions. The scalability challenge in those dimensions is manageability. The MS argument about scalability is basically that an enterprise can manage its IT assets more cheaply on NT (and manifestly looks easier to a PHB because you use the familiar windows GUI).
However, I think most people have figured out by now that the "user friendliness" of Windows is basically a cardboard facade put up on a big honking hunk of complexity. I think Linux (as well as other Unices) has the opposite problem in that a lot of its utilities appear unneccesarily complicated, but the underlying system is much cleaner and more modular. It would be cool if every package adopted the same scheme for its configuration files, perhaps XML based:
< netdef >
< netname > sales-dept-subnet <
< ipnetdef >
< ipnetaddr > 192.168.0.64 <
< ipnetbits > 26 <
<
These could be manipulated with any combination of GUI tools, Web tools, command line tools, or even special YACC grammars purpose built to your enterprise network.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
SGI: We are making a bid for the T30 supercomputer
CNET: What OS will you be using?
SGI: No comment.
CNET: (aha!) So what CPU will you be using?
SGI: No comment.
CNET: (Aha!) Can you confirm that you will be using Linux.
SGI: No comment.
CNET: AHA!!!!! (Whoops did I say that out loud?)
Later...
CNET: (Damn, I forgot to ask about alien technology, I guess I'll just go with the Linux angle. I'll throw some "experts" in there for "balance". Slashdot readers will read it....I'll be rich) Rich I tell you!! haHaHAHAHAHAHAH!!!!!! (cough)
--
"L'IT c'est moi!"