IP: Cringely's Supercomputer

[David Farber <dave () farber net>]

> From: Dewayne Hendricks <dewayne () warpspeed com>
>
> [Note:  This item comes from reader Steve Stroh.  Hummh... like his 
> earlier columns on 802.11b, Cringely gets a number of things wrong in his 
> description of UWB.  Yet again we get an explanation of the technology 
> that borders on magic from someone who writes for the media.  DLH]
>
> At 21:06 -0800 12/26/01, Steve Stroh wrote:
> > From: "Steve Stroh" <steve () strohpub com>
> > To: "Dewayne Hendricks" <dewayne () warpspeed com>
> > Subject: Cringely's Supercomputer
> > Date: Wed, 26 Dec 2001 21:06:55 -0800
> > MIME-Version: 1.0
> >
> >
> > Interesting wireless twist at the end!
> >
> > Thanks,
> >
> > Steve
> >
> > --
> >
> > DECEMBER 27, 2001
> >
> > What I Want for Christmas
> > A Supercomputer in Every Garage!
> >
> > By Robert X. Cringely
> >
> > Reality has lately been a little too REAL for me. The economy is tanking,
> > we're at war, the national and international situations simply cry out for
> > escape, denial, and delusion. Why worry when you can nerd out, instead?
> > That's when I decided that what I really wanted for Christmas was my very
> > own supercomputer. Doesn't everyone?
> >
> > Ignoring for the moment the pressing question of why any individual would
> > actually need a supercomputer, I prefer to revel in the fact that it is even
> > possible to have one. We live at a time when processor and memory prices are
> > at historic lows allowing enormous amounts of computing power to be
> > accumulated in the back bedrooms of houses like mine. The minute I realized
> > that I could have a supercomputer, I had to have one. But would Santa come
> > through? Given all the poor children in the world far more deserving of
> > gifts, I decided to take the obligation from Santa's shoulders and build the
> > supercomputer myself.
> >
> > Building a supercomputer these days is pretty much a matter of throwing a
> > lot of processors and memory into a big box or boxes, then finding some
> > cheap clustering OS (usually Linux) to make it all work together. My role
> > model in this venture is KLAT2, the Kentucky Linux Athlon Testbed 2, which
> > in a recent ranking came up as the 200th most powerful supercomputer on the
> > planet. I love the idea of a supercomputer from Kentucky, but I love even
> > more the clever design of KLAT2, which is a cluster of 64 PCs each running a
> > 700-MHz Athlon processor for a total of more than 64 gigaflops of
> > number-crunching power. Built by University of Kentucky graduate students
> > led by professor Hank Dietz, KLAT2 cost only $41,000, which is cheap for a
> > supercomputer.
> >
> > What's so clever about KLAT2 is the way the 64 separate PCs are linked
> > together. This is the real bottleneck in building a high performance
> > computer at low cost. Processors are cheap, memory is cheap, disk storage is
> > cheap, but networking -- at least really fast networking -- is still
> > expensive. To build a network capable of keeping up with those 64 Athlons,
> > the obvious choice would have been to use gigabit Ethernet adapter cards.
> > But gigabit Ethernet cards are expensive, and at the time would have cost
> > more than the PCs in which they were installed. There had to be a cheaper
> > way of connecting all those PCs together.
> >
> > So Dietz came up with a whole new network topology that allows cheaper,
> > slower network cards to perform as well or better than gigabit Ethernet. The
> > solution was to put more cheap Ethernet cards in each PC, and then use
> > "channel bonding" to make them all look like a single faster card. Dietz put
> > four 100 megabit-per-second fast Ethernet cards in each PC. Each card has
> > only one tenth the speed of a gigabit card, but not even gigabit Ethernet
> > cards actually carry a billion bits per second. With channel bonding, it
> > turns out that even using three cheap network cards per PC allows greater
> > throughput than a single gigabit card. And fast Ethernet (10base-100) costs
> > about three percent of gigabit Ethernet on a per-card basis, so using four
> > cards per PC still saves 88 percent.
> >
> > It is easy to use channel bonding for higher performance when the number of
> > machines is low and they can all connect through the same switch, but when
> > the number of processors goes up in the dozens or hundreds, the network
> > topology requires a lot of calculating. This is called a Flat Neighborhood
> > Network, and Dietz had to devise a genetic algorithm to calculate the switch
> > configuration that made for the lowest latency connections.
> >
> > The other KLAT2 advance is using the Athlon processors' 3Dnow! instructions
> > to bring parallel execution inside each processor, not just to the Linux
> > cluster. 3Dnow! is very similar to Intel's Multimedia Extension (MMX)
> > instructions and can greatly speed up certain calculations. Using 3Dnow! in
> > the KLAT2 required some C compiler changes that led to an across-the-board
> > 3X speedup. Now I wanted to apply those KLAT2 design features to my little
> > supercomputer.
> >
> > Fortunately, I wouldn't be starting from scratch. In my office, I had the
> > bones of an old Internet business venture gone sour -- six rackmount
> > computers. They used antiquated Cyrix-233 processors, but I mainly valued
> > the systems for their cases, fast Ethernet cards, and 9.1 gigabyte, 10,000
> > RPM Seagate Cheetah SCSI hard drives. I'd be replacing the ATX motherboards
> > and adding network cards in every case. I would need six new motherboards
> > and a lot of memory.
> >
> > The process is moving forward only as I can afford it, but so far I have six
> > new dual Athlon XP motherboards, each with a pair of 1.4-GHz processors, a
> > gigabyte of DDR RAM for each machine, and a total of 24 network adapter
> > cards. These will all plug into a 24-port 10/100 network switch to create a
> > flat network neighborhood. While my supercomputer won't be quite as fast as
> > KLAT2, it will be a lot smaller and cheaper at right around 24 gigaflops,
> > which is comparable to a top-of-the-line Cray T90 supercomputer from 1998.
> > And while that Cray cost millions, my out-of-pocket supercomputer budget is
> > $6,000.
> >
> > Unlike KLAT2, the operating system for my little supercomputer won't be
> > Linux. It will be QNX, a real time OS that supports massive parallelism and
> > has very low overhead. QNX is fast! QNX is also Posix compliant, so there is
> > lots of software that almost works under it. And even though QNX is a
> > commercial operating system, it is free for noncommercial purposes like
> > mine.
> >
> > Beyond using it to heat my office, I plan to keep the supercomputer busy
> > with a video compression project I'm doing as well as further experiments in
> > wireless communication. Having not learned any painful lessons from my long
> > distance 802.11b experiments, I've decided to get even wackier in my
> > attempts to improve Internet connectivity and will be looking into Ultra
> > Wide Band networking. UWB is a form of wireless data communication that uses
> > radio in a completely different way, sending short pulses of energy across
> > the entire zero to 60 GHZ frequency band. Not long ago, only spies and
> > Secret Service agents used this stuff, but now there are many companies,
> > including Intel, that are developing UWB chipsets. UWB could replace
> > communications of all types, ending forever our dependence on wires and
> > making worthless the ownership of radio frequencies.
> >
> > UWB is like magic or quantum mechanics, whichever you prefer. It is immune
> > to interference just as it doesn't interfere with traditional radio signals,
> > so the FCC is considering UWB as an unlicensed service across all frequency
> > bands -- even cellphones and broadcast frequencies. How could they regulate
> > communications they can't even detect? UWB uses one ten thousandth the
> > energy of networks like 802.11b, yet offers the prospect of greater range
> > and greater privacy along with data rates that are presently around 60
> > megabits-per-second and might eventually hit one gigabit-per-second. UWB is
> > virtually undetectable by traditional radios, since its signals are
> > considered noise -- noise spread across such a wide band as to be beneath
> > the threshold of traditional receivers. UWB uses multipath interference as a
> > form of error correction! What was formerly considered bad is now good. In
> > fact, UWB only works at all because we know precisely where and when to
> > listen. It is based on a complex and very rigidly-structured encoding
> > scheme, and that's where my little supercomputer comes in.
> >
> > I've decided to call her Wendy.

For archives see:
http://www.interesting-people.org/archives/interesting-people/