Linux Desktop Clustering - Pick Your Pricerange

crashlight writes: "A Linux cluster on the desktop--Rocket Calc just announced their 8-processor "personal" cluster in a mid-tower-sized box. Starting at $4500, you get 8 Celeron 800MHz processors, each with 256MB RAM and a 100Mbps ethernet connection. The box also has an integrated 100Mbps switch. Plus it's sexy." Perhaps less sexy, but for a lot less money, you can also run a cluster of Linux (virtual) machines on your desktop on middle-of-the-road hardware. See this followup on Grant Gross's recent piece on Virtual Machines over at Newsforge.

Comments

Virtual macines???

[TheAwfulTruth]

The purpose of running clusters is to increase processing power and/or fail-safety. How is running 8 virtual machines in any way a "less sexy" version of an 8 CPU cluster?

Re:Virtual macines???

[PD]

Maybe I want to develop software for a Beowulf cluster, but either I don't have a cluster of my own, or I just want to write the software, not run it in production. Either way, a set of 8 virtual processors would be good enough for the job.

Re:Virtual macines???

[PD]

I just thought of something else. I have never used a Beowulf cluster, so maybe I'm completely wrong, but virtual machines could make a Beowulf more easily upgradeable. The idea is that you'd make a cluster with a whole bunch of virtual machines, say 1024. The cluster is fixed at that size for all the software that runs. But in reality, you've got 32 processors actually running. When you upgrade the cluster to 64, you don't need to reconfigure any of the software that runs on the cluster, because they all assume that you've got 1024 processors. But, you get a performance increase because there's now more physical processors. As I said before, I don't know much about clusters. I imagine that somebody who really does know will quickly either confirm what I said or reduce my idea to a pile of stinking rubble.

Re:Virtual macines???

[cweber]

You're mostly off the mark, I'm afraid. Most software that uses a cluster runs through MPI or simply through scripts. Both mechanisms allow for easy adjustment in the number of nodes/CPUs you use.

Many large compute problems are embarassingly parallel, i.e. the same calculation needs to be repeated with slightly different input parameters. There's basically no interprocess communication, just a little forethought about filenaming conventions, total disk and memory usage, etc.
Execution of such tasks reduces essentially to a simple loop:
foreach parameter_set
rsh nodeN myprog outfileM
end

For those programs that actually run a single instance of the code on several CPUs, you have to be acutely aware of how many nodes you use. Your code has its own limits on how well it scales to multiple CPUs, and your cluster imposes limits on how well (in terms of latency and bandwidth) nodes can communicate. Very few codes in this world scale well beyond 64 CPUs, especially not on run-of-the-mill clusters with plain ethernet interconnects. Fortunately, it is trivial to readjust the number of nodes used for each invocation of the code.

Lastly, virtual nodes cannot easily simulate the behavior of real nodes. Again, it's the interconnect latency and bandwidth. When it comes to supercomputing, only trust what you have run and measured on a real life setup with your own code and input data.

Re:Virtual macines???

[Perdo]

A beowulf cluster operating system could expose a standerdized set of API's to an application if no matter how many processors it had it told the application that it had 1024 but the application could not then be optimized for the actual hardware it is running on. KLAT2's applications were optimized to take advantage of the 3DNow! extentions which Trippled the effective processing power of their cluster. Just like Direct X provides API's to applications for standardization does not come close to what could be achieved hand codeing machine language to the x86 hardware. Look at the applications being run on clusters now and you will see that most are cash poor but manpower and entusiasm heavy. If they could afford 3 times as much hardware to achieve the same performance they would have ordered expencive big iron from IBM. Also not that 1024 processors limits the scalability of the cluster. The real Idea is to force hundreds of commodity machines to act like a single processor

Re:Virtual macines???

[rew]

The idea is that you'd make a cluster with a whole bunch of virtual machines, say 1024.

Good plan. There's a large BUT however....

There are two sources for N processors not being N times as fast as one.

First there is the communications overhead. If you have 1024 virtual processors, running on 32 real processors, there are going to be about 32 programs running and communicating on one processor, while they could've gotten by without communicating at all...

Also, the problem may not be divisable in 1024 equal chunks.

If you write your software to scale to 32, there is going to be a parameter that says how many processors there are already. So adjusting it to 64 is not going to be hard.

But test-running on a small (but virtually large!) cluster before running it on the "real cluster" could allow you to guess the magnitude of both "performance impacting issues" before buying time on the expensive "real cluster". So there certainly is a point in doing it that way....

Roger.

- 8 - 800MHz Celerons?

[qurob]

So, it's about equal to my dual Athlon 2000+?

Wow... what a markup!

[eaglej]

Yeah, it's a nice compact little box... But they're pulling in a phat few g's on each box. I'll build my own, thank you very much.

Why is Timothy badmouthing...

[SpookComix]

...the articles he posts? Tim, baby, if you don't like or agree with the article, either don't post it or keep your opinion to yourself.

I know I'll get modded down for this, but here's an example:

Posted by timothy on Tuesday January 22, @02:45PM
JackBlack tells us about the "unbelievable deal you can get at KMart on all their overstocked computers and periphials! You won't believe the kind of prices on these things! I don't know about you people, but I'd rather swallow Draino than buy boxen from Big K! But, shit, whatever, I guess.

Timothy is his own conflict of interest.

What is happening to Slashdot these days?

--SC

Re:Virtual vs physical

[f00zbll]

I don't know much about building a cluster, but I thought one of the limiting factors of clusters is network latency. If they have it all on the same motherboard, or designed their own bridge to connect multiple CPU/MB, it would be faster right?

By the fact that they are all incased in the same box, rather than connected via a switch, it has less distance to travel. I don't know that 5 feet of CAT5 could make that big of a difference. On the otherhand, they could have designed a different way of bridging the systems and dramatically reduce latency. In either case, it is intriguing.

under engineered.

[azephrahel]

I'm sorry, but for that price this is way under engineered. The origonal bewulf cluster was made with components on par, for the day, as the celeron modle of the redstone, for far less. If your going to spend the time and money building and marketing systems like this, they could have done a better job. They suck mobos in a big case and eth linked them togeather. Call me crazy but I think for that much money you could get a small backplane, 8 industrial PC's (powerpc/copermine/whatever on a pci card each w/its own memory) toss em in and spend the rest of your "engineering" budget, making a patch to the kernel for reliable communication over the bus, instead of slow eth connections.

besides with the speed advantages shared memory brings to multiprocessing a quad xenon would probably outpreform this.. deffinately a quad proc ultrasparc but those are pricey even used...

Re:under engineered.

[benedict]

Sure, you could build it yourself, and your
hardware cost would be lower.

But what about the cost of your time, in terms of
dealing with vendors, putting machines together,
testing them, integrating them, and testing again?

I'm guessing these machines come with support, too,
though I can't tell because their web site is
Slashdotted.

Re:under engineered.

[mpe]

If your going to spend the time and money building and marketing systems like this, they could have done a better job. They suck mobos in a big case and eth linked them togeather.

But for some reason chose a midi tower, rather than a 4U rackmount.

Re:under engineered.

[zerocool^]

besides with the speed advantages shared memory brings to multiprocessing a quad xenon would probably outpreform this.. deffinately a quad proc ultrasparc but those are pricey even used...


First, I'm guessing you meant quad-Xeon and not quad-Xenon, cause why would you want to have 4 instances of Xenon text editor running on your desktop, or 4 atoms of the element Xenon.

Second, there is the misnomer that Ultra-sparc boxen are rediculously expensive. It's just not true. The place where I work recently purchased a new shared solaris server - quad-proc Ultra2 4x300mhz w/ 1 Gig of ram and 2x18 gig scsi drives. It only set us back about $1500. That is NOT MUCH in the world of shared website hosting. Plus, it compiles things a lot faster than, say, our P4-1.5 Ghz, despite the 300 Mhz advantage.
High end hardware is doable, and most people don't need much, especially considering that a Sparc IPC (25-ish mhz w/ 24 megs of ram) runnig solaris 7 can host 200 web pages easily, and handle 3 million hits a month. Web serving is *not* difficult, and it doesn't take a lot of power, just a proc that can switch between processes quickly and efficiently. It's poorly-written CGI that takes power.

~z

Re:under engineered.

[zerocool^]

(censored)> su -
Password:
Sun Microsystems Inc. SunOS 5.7 Generic October 1998
You have new mail.
(censored):root> mpstat
CPU minf mjf xcal intr ithr csw icsw migr smtx srw syscl usr sys wt idl
0 16 0 86 45 44 70 0 1 2 0 98 1 1 1 97
1 16 0 88 1 0 72 0 1 2 0 102 1 1 1 96
2 16 0 115 213 13 67 0 1 2 0 100 2 1 1 96
3 16 0 95 67 67 64 0 1 2 0 97 2 2 1 96

Tell me how to get more detailed system info and i will, i don't know how to just print out exactly what the procs are and speed ratings

Re:under engineered.

[zerocool^]

Hey, how's this?

(censored):root> dmesg

Jan 22 21:47
cpu0: SUNW,UltraSPARC-II (upaid 0 impl 0x11 ver 0x20 clock 296 MHz)
cpu1: SUNW,UltraSPARC-II (upaid 1 impl 0x11 ver 0x20 clock 296 MHz)
cpu2: SUNW,UltraSPARC-II (upaid 2 impl 0x11 ver 0x20 clock 296 MHz)
cpu3: SUNW,UltraSPARC-II (upaid 3 impl 0x11 ver 0x20 clock 296 MHz)
SunOS Release 5.7 Version Generic 64-bit [UNIX(R) System V Release 4.0]
Copyright (c) 1983-1998, Sun Microsystems, Inc.
mem = 1048576K (0x40000000)
avail mem = 1029668864
Ethernet address = (censored)
root nexus = SPARCengine(tm)Ultra(tm) CycleQUAD (4 X UltraSPARC-II 296MHz)
pci0 at root: UPA 0x1f 0x4000
pci0 is /pci@1f,4000
pci1 at root: UPA 0x1f 0x2000
pci1 is /pci@1f,2000
etc.etc.
I rest my case. BAM.
~z

Re:under engineered.

[zerocool^]

eh, it's not new. I mean, we bought it used, and it's a couple of years old. However, we did pay like $1560 for it, plus shipping. I'm not sure where from, and i can find out if you'd like, but I'm still amazed by the price. Which is why i wanted to post it. People looking for high end systems(64 bit) w/ quad procs can get them for a reasonable price.

~z

Rack Density

[Genady]

So... how many processors can you fit into a standard 44U enclosure now? If they've got an integral Ethernet switch do you get a gigabit uplink out? This would actually be really cool for Universities/Government Agencies to build insanely great clusters with small floor space. Still if you want insanely great maybe you should cluster a few briQ's together.

Re:Is there a real market for this?

[rhdwdg]

I could. The form factor is the thing. I could use a few extra CPUs in a MOSIX cluster for my desktop, but I have no room for a small rack and associated power. This fits. I could make them into little application clusters -- 256 MB of flash is plenty per device. I could wish they had GigE, of course (since they obviously need to connect to NAS for data) or multiple NICs per system but even 100 Mb is sufficient for the intended markets.

Terrasoft, briQs, PPC, 4-8 nodes...

[tcyun]

I saw a quick demo of a multi-noded briQ (by Terrasoft Solutions) at SC2001 a few months ago and was very impressed. The ability to leverage the power of the PPC in vast numbers (and in a very small form factor) was incredible. I wonder how these would do in a head to head competition?

They offer a 4-8 node tower running 500 MHz G3 or G4 CPUs and drawing "roughly 240 watts per 8 nodes (less than a dual-processor Pentium-based system)." Quite impressive.

Re:Terrasoft, briQs, PPC, 4-8 nodes...

[gonar]

try $1500 for a g3/500/512mb node
and $1500 for the case

so for $4500 you get 2 count 'em 2 g3 nodes (no altivec)

$2k for the g4 node (same speed & mem)

so for an 8 node cluster, you are looking at $17,500

makes the rocketcalc machine look like a bargain, no?

(ok, ok, the briq is better engineered, but still...)

Here's what I'd much rather see

[S. Allen]

I'd like to have see chips that incorporate the CPU, RAM and something equivalent to the North+South bridge. Motherboards should be designed to take 1-32 of these plugged into some godawfulfast bus. CPU and RAM should be one in the same and scale together. RAM co-located w/ the CPU would be much, much faster. Most systems and applications can scale with more threads or CPU's. CPU's by themselves are just about as fast as they need to be for any task that cannot be divided into multiple threads (I'm not talking about poorly written progams). This whole getup would be significantly more elegant, reduce parts and complexity and probably be cheaper to produce in the long-run.

I don't see this as the same as a system-on-a-chip. With those, you're integrating video and audio. I'd either rather NOT see that integrated at all or have a portion of this new CPU combo thingy incorporate a DSP or FPGA region(s).

Whoa, time to put down the crack pipe.

Re:Here's what I'd much rather see

[mpe]

I'd like to have see chips that incorporate the CPU, RAM and something equivalent to the North+South bridge. Motherboards should be designed to take 1-32 of these plugged into some godawfulfast bus. CPU and RAM should be one in the same and scale together. RAM co-located w/ the CPU would be much, much faster.

Most of this (apart from the RAM with the CPU) sounds like a Sequent Symetry. There's also the Vax cluster. Where processors (with RAM) and storage controllers connect to a 16 (IIRC) way star interconnect.
Both of these are over a decade old technology. So something similar which fits in a box one person can pick up is probably overdue.

only 100mbps?

[Restil]

The primary disadvantage of clustering is the network bottleneck. You lose out because even 100mbps is only a small fraction of what the pci bus of even low end pentium systems are able to handle. At LEAST go with gigabit ethernet so you can push over 100 megs per second between processors. This will greatly increase the usefulness of an integrated cluster by decreasing the one primary disadvantage.

Also a bit pricey, but there would be some cost advantage in reduced footprint for some environments.

-Restil

Re:only 100mbps?

[Lumpy]

being in the same box I wonder why they didn't use U160 scsi for comms. hell you can have 16 nodes on a SCSI bus nowdays and transferring data between 8 machines and the array of shared drives would make it scream like a raped ape.

hell we use that kind of setup for spitting 24 mpeg2 streams out for commercial insertion from pentium 166's

Re:only 100mbps?

[mpe]

The primary disadvantage of clustering is the network bottleneck. You lose out because even 100mbps is only a small fraction of what the pci bus of even low end pentium systems are able to handle. At LEAST go with gigabit ethernet so you can push over 100 megs per second between processors. This will greatly increase the usefulness of an integrated cluster by decreasing the one primary disadvantage

Depends on the application in question. There are many parallel processing tasks which do not need massive communication between processors. Effectivly each processor simply gets on with it's task on it's own.

Re:only 100mbps?

[GigsVT]

Do you have any information on setting up a shared SCSI bus like this? I'd be very interested in doing it, but I have not been able to find any information about it.

gigs at vt dot edu

Re:only 100mbps?

[Lumpy]

Yes start here

http://linux-ha.org/comm/

this is used in the linux HA system extensively. Xenix and Unix has it also.. No version of NT has this ability nor any Microsoft product. Maybe someone has written a driver for NT4.0 but I doubt it.

Re:only 100mbps?

[sporty]

Took a bit of thought to what you said. Wouldn't you be communicating smaller pieces of information on the norm vs huge amounts? Or are you tinking of doing large data translations? Perhaps they were building for one and not the other?

Heh, there'll always be a bottle neck.

Re:only 100mbps?

[benwb]

100mbps ethernet consumes 75% of a 32 bit/33 mhz pci bus. Jumping up to 64bit/66 mhz it's still consuming nearly 20% of available bandwidth. The current generation of PCI busses (even "high end" ones) cannot pump out enough information to saturate a gigabit line.

Re:only 100mbps?

[Wesley Felter]

That's totally bogus. PCI is 32 bits * 33MHz = 1056 Mbps.

Re:only 100mbps?

[SpinyNorman]

Nice in theory, but in practice Gigabit ethernet does not deliver Gb/sec - and in fact Ethernet's contention access mechanism (CSMA/CD) guarantees that you won't!. In practice you're better off with channel bonded 100Mb/sec Ethernet and a smart switch topology such as KLAT2's Flat Neighborhood Networks to minimize latency and maximize inter-node bandwidth.

KLAT2 FNN

Note to moderators: above link refers to Linux, AMD processors and Beowulf clusters! Do the right thing! ;-)

Re:only 100mbps?

[wadetemp]

Uh huh... have you ever used a reasonably new 32bit PCI computer when it's transfering data at close to 100mbps? Let alone when it's actually supposed to do something with that data, like write it to a hard drive or process it. Want to multiply that effect by 2... or 3... or 8... or 10?

Re:only 100mbps?

[Drakantus]

Yes I have. Copying files from one drive to another over the PCI bus often transfers at 15-25MB per second, which works out to 120-200mb/sec. The CPU shows some usage and things are a little sluggish, but the OS is still totally usable, especially if it's on a 3rd drive that is not involved in the copying.

Besides, due to inefficiency in TCP/IP you can't really get the full 100mb/sec, it's mostly limited around 80mb/sec for the better NICs.

Virtual clusters?

[Lxy]

you can also run a cluster of Linux (virtual) machines on your desktop

So you're suggesting that I divide my machine into 8 virtual machines and then cluster them for uber fun? Figuring the extra latency, wouldn't it be faster to just leave it alone?

Re:Virtual clusters?

[athakur999]

The point isn't gaining speed, it's gaining reliability and maintainability. Instead of one PC being your email, database, proxy, etc. server, you create a virtual machine for each one.

If you need to upgrade the kernel on your email server, you can just reboot that virtual machine and leave your database, proxy, etc. servers untouched. If a process goes wild on your database server it'll only screw with that service and not your email, proxy, etc.

Clustering under VM is....pretty dumb

[CDWert]

Clustering under vserver scheme is pretty dumb, if it was in fact meant to be serious.

vservers assume that a machine has resources avaiable and that no one instance is consuming 100% of the systems resources. Application built for utilization under a cluster would most likey CONSUME 100% quickly and easily, otherwise why run them under a cluster in the first place ????

I do like their monitor applet, its pretty coll for basic cluster managment/monitoring.

At the same time all you people complaining about price...lets see

800mhz cele w256 meg each MB, NIC and whatever storage, lets say on the cheap $300 each

4 350 Watt PS $100

Ok a really cool case and PS

Time to load software, lets say weve done it before and it take 8 hours

Physical Assembly and testing 5 hours

My bill rate and personal time is worth 120/hr
$1560
Hardware $2400
Power Supply $ 100
Mildy cool case$ 100

THAT COMES TO 4160, Hell, add to that I dont have to build the SOB AND it comes under warranty , you betch you A** Id buy one of these

PLEASE think before you gripe about prices...
Looks like a deal to me ....

PS, Kernel version is 2.4.12 from what I saw on their link to products showing a screen shot Ayeee.....

The Needed Ingredient

[4of12]

OK, the price for these kinds of things is really nice and low. Low enough to make anyone in the numerically-intensive computing arena really want something like this.

(However, I probably wouldn't want one of these as my desktop machine if the power supplies took more current than a typical wall outlet, if it made as much noise as a helicopter taking off, and if heated up my cubicle to 92 F.)

But the key ingredient in my mind is making these distributed boxes more conveniently usable, much like those 64-way big boxes from Sun and SGI.

How far along are MOSIX, Scyld's products, others(?) that make these distributed clusters have a nice Single System Image?

Re:Hey

[ch-chuck]

Sure, why not. If an 'internet' is a network of networks, we should be able to build a cluster of clusters. One cluster calculates the reality I'm flying thru while another one calculates the effects of the nuclear device I just heaved, both feed into the headmounted 3D stereo graphics processor with surround sound audio helmet on the hydraulically actuated platform, while an input/output cluster handles sim data from the other players over the fibre channel...

Do-it-Yourself Cheap Clusters

[ChaoticCoyote]

I've been building my cluster from various remaindered/cast-off/refurbished machines I find. Computer Geeks is a good source.

Load balancing is frelling difficult, but I've been doing some solid parallel programming work that translates nicely to "real" clusters. I'd love to buy one of the Rocket Calc boxes -- but I can make a darned nice box for a lot less money with more processing power, if I'm willing to have cables and such all over the place.

The only real cluster-related problem I have is my lovely wife. She's one of those people who want things to "match" (so why in frell did she marry me?), and my "heterogenous" cluster just isn't very aesthetic. She just doesn't understand that the cluster's job is to compute, not to look pretty!

Then again, the Rocket Calc machines are attractive, and the color would go with the living room furniture...

Re:Do-it-Yourself Cheap Clusters

[Courageous]

frell? frelling???

You mean we weren't supposed to all understand that you really meant "FUCK" when you said that? Try it. Repeat after me. FUCK. F-U-C-K. FUUUUUUCK! Feels good, doesn't? Wheh, now aren't you glad you've got that out of your system?

I knew you would be.

>:-/

C//

site slashdotted, here's google cache

[eufaula]

here is the google cache since the site seems to be slashdotted....

http://www.google.com/search?q=cache:q2egeLKmoVQC: www.rocketcalc.com/+rocketcalc&hl=en

Re: - 8 - 800MHz Celerons?

[qurob]

Although I'm joking, let's just take a look at some numbers, hypothetically speaking.

*borrowed from Tom's Hardware*

Linux Compiling Test

3.35 minutes for a Athlon XP 2000+
14.2 minutes for a Intel Celeron 800mhz

(now, here's where we stretch it)

Figure 1.7 minutes for a dual Athlon XP 2000+, 50% of the other time.

1.7 x 8 = 13.6 minutes


But, who really compiles with a cluster, really?

It'd still be faster....At least on a few benchmarks, and at least in theory

Two Questions

[fm6]

(1) Since this is all about high-performance computing, why use Celerons?

(2) Why is it taking so long for someone to make the obligatory "Imagine..." post?

whatever.

[jonbrewer]

To run multiple linux instances on one "middle of the road" server, you need VMWare GSX. It ain't free. In fact, it's $3,550.00. (there goes your "a lot less money" idea, T.)

As for the value of this product, I see it clearly. Not all computational problems need high data throughput between nodes. And their Redstone-A product gives you an 8 node PIII 1Ghz cluster with 4GB of ram for $6000. And all the networking set up and ready to go. Give it to your Scientist and they don't need to know jack about network or configuration, they just treat it like another unix workstation.

When I think of the ~ $20k each we spend on Sun and SGI workstations for our scientists, I cringe. This I wouldn't (won't?) think twice about buying.

a desktop renderfarm

[Lumpy]

You know when spielberg said that the next big smash wont be from him or some big studio but it will be from a woman named buffy from idaho. This brings that possibility one step closer. I can see this as a renderman render farm or even just an effects renderfarm for some of the effects the AVID generates.

gimmie 2 of these, an AVID, and a $200,000.00 camera and lens you've got the next ET on your hands... (Nooo not cheech and chongs Extra Testicle get your mind out of the gutter)

it's getting there where the general schmuck has the power that hollywood does.

Re:a desktop renderfarm

[Graymalkin]

Talent != equipment. You can do a film is craptastic 8mm with butcher paper soaked in crisco as a lens and still make a rad movie if you've got a smidgen of talent behind the camera and it helps to have some in front of the camera too. I'd rather have a real 8-way system than a cluster-in-a-box. The 8-way with the right software is going to be alot more productive than an 8-node cluster limited to batch rendering stuff I throw at it. Shit I'd rather have four dual processor systems with fast storage systems and 4 talented guys on them rather than 8 processors and one talented guy.

Re:Linux not ready for the desktop [was Re:Expensi

[r00tdenied]

Cost of upgrading what? Did you even read the article? This is a CLUSTER, not your run-of-the-mill desktop or workstation. I could get linux to easily run on an old 486 motherboard that is somewhere in the bottom of my closet.

If any OS is expensive due to upgrades, it is definately Micro$oft OS's. Can you see Windows XP running on a 486 33 mhz? I thought not.

Additionally linux cost a LOT less to administer by IT shops than Microsoft operating systems. With Microsoft operating systems, you have to *click* here, *click* in this text field,etc. I could have ipchains up and running fast than you could have NAT running on Windows2000.

However, this cluster is a great solution to a lot of problems. It would definately free up colocation rack space, and make it easier to do virtual hosting.

r00tdenied

Re:Cool software

[GigsVT]

Please never order from TigerDirect. Check the BBB record on them and resellerratings.com. They are the target of many consumer lawsuits, and have a terrible record.

and for the price . . .

[hawk]

$4500: 8xCeleron 800, 8x356mb=2g.

My new workstation (yes, it's finally coming, for those whohave been wondering; the purchase order goes to main campus today):
$4800: 2xAthlon 1900, 2gb ddr.

My memory and bus are significantly faster. I believemy total processing power is equal or close under most applications.

Then I get a few things that aren't in that bundle:
2x18gb cheetah 15000rpm u160
4x 9gb cheetah 15000rpm u160
21" sony monitor & video card
scsi cdrw

would I really get any more from this unit???

Re:and for the price . . .

[sigwinch]

The Celeron cluster also has 8X CPU cache. A workload that doesn't fit in dual Athlons' cache might fit in octal Celerons. Since main memory is ~10X slower than cache, the speedup in the inner loop can often pay for the slowdown from interprocessor communication.

Re:and for the price . . .

[hawk]

Ithas 8 L1 caches, but they're not 8 times as large. Come to think of it, how do the cache sizes compare? Doesn't each athlon have about 4 times the celery's, equalizing thimngs?

also, it seems more likely that one big cache wouldhave what you need than a collection of little ones *all* having it.

I also noticecd that I used 9 words for each array element in my stucture. *doh*. 8 would have exactly alligned it in terms of memory fetch operations (one or two; I forget)>

Anyway, isn't the 1900 athlon somewhere in the broad range of 3 800 celerys?

hawk

Re:and for the price . . .

[sigwinch]

Athlon XP data cache is 320 kB (ignoring storage of instructions in L2 cache). Dual machine has 640 kB total.

Celeron 800 data cache is 160 kB. Octal machine has 1280 kB total, twice that of the Athlon.

Even if Celeron cache transfer rate is half that of Athlon (don't have exact numbers), total throughput will still be double.

also, it seems more likely that one big cache wouldhave what you need than a collection of little ones *all* having it.

It depends on the workload. If you can parallelize the cache load, it can be a huge win. If all the caches are flushed by the same constant data that won't quite fit, it doesn't help much. As usual the problem and how you parallelize it are just as important as the hardware.

There's a better way...

[The Panther!]

If you're looking to do the same thing, cheaply and with a little DIY hardware tech time, simply buy a big ATX case with room for a couple of extra power supplies, install a 386 or 286 mobo (all ISA slots!!!), and buy a number of these computers and plug them in. They don't draw power from the motherboard, so you don't even bother connecting the 286 to power. Instead, tie each power supply to one or two of these cards directly (requires a little soldering), and there you go. A cluster in a box.

If I had the money, I'd be doing this myself. Instead, I've got a rack full of 4U AT cases with dual PPro 200mhz machines instead. The one advantage to having full sized motherboards (with PCI slots) is that I'm installing triple-channel-bonded ethernet so I get gigabit ethernet bandwith, without paying gigabit prices.

Thought for the day...

[jd]

The fastest processor out there (in terms of sheer power, not mere MHz) is the Athlon XP 2000+. I've checked with AMD, and they don't say you -can't- use it with the MP chipset. (They don't say you can, either. What they DO say is that they don't list or recommend that combination.)

That'd give you 2 processors per chunk. By stapling four chunks together, using MOSIX, you'd get the same as an 8-proc SMP box, without the hyper-expensive motherboard.

Now, this is where it gets fun. (Yes, I've been reading up a bit on this. :) The processor uses a clock multiplier, to translate the clock rate of the system into the clock rate for the processor. You -can- fiddle with this multiplier, but it occured to me that you don't actually need to. You can simply replace the system's clock, instead.

I don't know what the "default" multiplier is, but I'd guess it's probably x16, or thereabouts, given the speed of the rest of the system. In which case, if you threw in a 1GHz clock (I suggest oven-baked, as they taste better), you'd get 16 GHz on the processor.

Now, THAT kind of ramp-up is not going to be easy. Chips run hot as it its, and if you plan on overclocking an Athlon by a factor of 8, you can expect to fry eggs on the surface. As for the poor RAM chips.... those are REALLY going to suffer!

My guess is that if you strapped pelziers onto all the chips, immersed the entire system into some synthetic, non-conducting fluid that you can super-cool to, say, -135'C (there's some stuff 3M makes that'll handle it), and you devised some way to keep it that cold, the chips might survive the experience.

Might. I'm not even going to say "probably". To the best of my knowledge, no overclocking or supercooling experiment on conventional PCs has gone to that extreme. The only ones that came close (a NZ project involving pouring liquid nitrogen into the case) trashed the disk drives and BIOS.

On the other hand, I've been checking up on the tolerences of components, and what you COULD build, if money wasn't an issue. The technology for an 4x2 MOSIX/SMP cluster, overclocked to 16 GHz and still running -does- seem to exist. (The keyword is "seem" - most chip specs are calculated, not actually measured, according to the data sheets.)

Now, I suspect that it'd cost a damn sight more than $4000, just for the parts, even if you could mass-produce such a monster (assuming you had the engineering skill to build it in the first place), or even find anyone crazy enough to buy one, given that it'd probably be more cost-effective (and certainly safer) to go with an IBM mainframe at that point.

On the other hand, kernel compiles would be quick.

Re:Thought for the day...

[Knobby]

The fastest processor out there (in terms of sheer power, not mere MHz) is the Athlon XP 2000+.

I thought the IBM POWER4 processors were still out front..

Re:Thought for the day...

[ameoba]

One catch :

Anyone spending this kind of money on a cluster is likely to be doing serious number crunching. Now, some types of number crunching don't mind posibly letting a wrong answer slide through somewhere (IE - in a render-farm, you might get a bad pixel). Other environments, where calculations are all interdependant, a single bad number could make all the computation worthless (the old "A butterfly flaps its wings in Tokyo and we get another John Katz article" story).

Besides, no point in spending the time designing & testing this monster OCed box, when you can just buy a few 8-way Xeons, or call Sun/SGI.

Re:Sexy...?

[liquidsin]

I'm sorry but it's just a big red box.

What's not sexy about box?

Oh, you meant like a computer...I was thinking of something else...

Oh well.

[rice_burners_suck]

Oh well.

The other day, I saw an auction on eBay for an SGI Octane. The price was over 100,000 bucks. Looking at the pictures, I could tell why: It included a rack with a bunch of really fancy stuff on it, and an SGI Octane. That is why I consider sexy.

Actually, here is what I really want. I want to have several of every brand of computer, running all the operating systems available for each brand. That way, I'll be able to access software and information for any of them. It'll cost a ton of money, and that's money I don't have, but hey, who said you can't imagine stuff?!

Oh well.

If you need to get something out of your system...

[ChaoticCoyote]

...I suggest a laxative. ;)

"Frell" is a word used on the TV series Farscape; it has the nice ability to replace many different cuss words with on catch-all phrase. For example: "To frell with it!" or "Frell you." And using "frell" avoid those nasty negative moderations that can so bruise my tender ego.

As for "fuck", "hell", and other cursatives: I make sailors blush, youngster; I've been coding so long, I've had to invent or borrow new cuss words because I wore the old ones out. I'm bored with "fuck" the word "although not the act, mind you), and am looking for new, fresh alternatives.

Build server

[JohnZed]

I would love to use one of these (especially the higher-end model) for a build server. I seem to remember a parallel make utility for Solaris (one that allows you to use multiple machines to do a build). Anyone know of something similar for Linux?
--JRZ

Re:Linux not ready for the desktop [was Re:Expensi

[r00tdenied]

Again, not true. Casual users will be forced to read HowTo manuals and man pages [linuxdoc.org]. If you follow the link into the several pages, you'll see that some of them are *years* old!

Ipchains is not *years* old as you put it, so the howto's definately can't be. Many of the howto's are old, but that is becuase a lot of things are backwards compatible with the newer kernels.

r00tdenied

Re:Possible SGI killer?

[Graymalkin]

If you want to use a cluster to do the final rendering that might work but you can't use a cluster as a workstation. In order to have cache coherence and whatnot processors need a high speed low latency connection (latency being measured in nano seconds) not be connected by a NIC through the south bridge on the memory controller. If you were to stick 8 Celerons on a single node somehow then you'd give an Octane a run for its money.

Re:Hey

[LinuxHam]

If an 'internet' is a network of networks, we should be able to build a cluster of clusters

They do. It's called Grid computing. Now that Qwest can shoot 400Gbit/s over 1,000km or something insane like that, supercomputing centers are connecting clusters with faster links than your 64-bit PCI bus.

Claims about VMs

[zaqattack911]

Ok.. so VMs makes sense because it allows seperate virtual linux O/Ss for each major server function. (web database ftp..). This is good for management and accounts. However, "they" are still making claims that it offers somekind of performance advantage over running them all on one O/S. I mean... call me crazy but I don't see how there is much of a performance advantage.. if not just wasted memory here. Someone please throw me a clue.

this machine is nice.

[Billly Gates]

I can play solitaire and code so much faster with this baby. Come bring it on.

Marketing personal clusters

[heroine]

I would have been ecstatic if someone made personal clusters consisting of dual 1.5Ghz Athlon nodes on gigabit ethernet a long time ago yet only now has someone decided to market personal clusters of single Celerons. Time to give up on the technically possible and base wish lists only on the technically marketable.