Slashdot Mirror
The Ultimate Linux Box 2001
savaget points to this Linux Journal article which covers building a superior personal computer for general usage. See if you agree with the choices that Rick Moen, Daryll Strauss and Eric Raymond made in building their dream box.
My budget doesn't allow ultimate boxen... I'd be more interesting in seeing information on ultra-cheap (but still decent and reliable) systems. An older guide exists, but it hasn't been updated in a long time.
Incase you are wondering why he is building two, it clearly states that one is for him, and the other is for Linus Torvalds!
Lucky dog.
This artical lost me when they all ignore IDE.
They and a lot of other people superficially judge SCSI to be faster than IDE.
If one has $1000 to spend on storage they you will get more bytes and better throughput with an IDE raid than a SCSI raid.
Why compare drives on a unit v's unit basis, there is no reason a computer cant have more than one.
Granted software raid is pretty crappy in linux, it doesnt scale well, but still, 2 7200RPM IDE disks in raid 0 will out perform a 10000RPM SCSI drive, be cheaper and provide more storage.
SCSI is for ignorant people with more money than sense.
Exactly how do you cluster IDE drives??? With SCSI I can share the the same bus with 2 different computers, and can present the same disk to two different systems at the same time.
o n1 1
a sc si.pdf
i to rial.html?prodkey=io_comparison
9 /
IDE is *only* good in a single drive / single controller situation; but at that time (from most drive manufacturers websites) you are only able to push maybe 35MB/sec. So your so called controller latency is NOT an issue. Agreed IDE will perform the same on a single drive system, but as soon as you add another drive onto that channel you've possibly halfed the performance of those two drives, you could add another controller, but really starts getting rediculus (I've got one systems with over 300 drives connected to it, I'd like to see an IDE system keep up with that)
There also are quite a bit of things in the SCSI protocol that you are looking over. Command Tag Queueing is a very big one, I can send multiple commands down the SCSI chain and the drive can re-order them so that the drive can streamline where it's going to be getting data off of the drive (setting this gives a significant performance boost on our arrays). Along with the fact that IDE is completely and totaly CPU driven, try really pushing your CPU and you are either going to have to give up CPU cycles to your app or give up performance to your drive.
Could you please provide a link to Google's use of IDE drives for all their storage, I can't seem to find a page saying that their Linux are all running on IDE only.
http://www.acc.umu.se/~sagge/scsi_ide/#comparis
http://www.dell.com/downloads/global/vectors/at
http://www.adaptec.com/worldwide/product/marked
http://www4.tomshardware.com/storage/01q1/01012
I don't understand how one can call any box without hardware RAID an Ultimate Linux Box. Yes, it costs money (which is why I've decided on IDE RAID for new systems unless I miraculously get about 20k USD for storage subsystem), but it's better.
Never go for a non-redundant disk subsystem. Disks crash. Go bad. Die.
Also, I need some space to live with. About 200 gigs minimum for my next setup.
So, for ULB it would be a dual channel U160 RAID controller (64bit PCI bus, please), 14 36GB 10k rpm hotswap disks configured as 6 disk RAID5 with hotspare on each channel and mirror over channels, yielding 180GB. Takes one SCSI tower case. Performance and redundancy. And even for the ULB going for 14 72GB disks would be pretty expensive.
Also, to get the best out of CD-ROM/R/RW get Plextor UltraPlex 40 and PlexWriter. Absolutely forget all other CD-ROM/R/RW manufacturers.
Then for DVD get the Pioneer and a DVD-R(G), DVD-RW drive. Note that You still need the Plextors to get the best CD-ROM/R/RW drives available.
Hook all four into an external case, and put the DDS-4 drive there, too.
Now, put the computer case and the drive case away from Your table, take the removable media case onto Your table and be done with it.
I can't afford that setup. So, I'm going for IDE drives and 3ware Escalade IDE RAID controller. Cheaper, and gives me about 240 gigs with seven 80GB drives.
2) I'm going on secondhand information, but I've generally heard better things about IBM drives than Seagate ones.
3) Use the Hoontech Digital 4DWave NX. A hardware manufacturer that actually *supports* Linux, releases all the specs on their hardware, has really good Linux drivers (look at ALSA), hardware mixing...
5) Logitech also makes a line of excellent optical mice, if like me you dislike MS products. There's this really neat one that has an intense blue light-up logo on top...
6) Mushkin's my favorite memory vendor. I've seen sub-par RAM be a PITA to diagnose twice now...top of the line RAM is worth the few extra dollars. Unless you've got an older system, getting marginal memory that doesn't *quite* work with your new motherboard is a real risk.
8) Arctic Silver II? You really *don't* need it
unless you're overclocking. It might make a degree of difference...but, to quote Half-Life, it's "well within acceptable limits" running without thermal compound.
Also....
*) Use *3com* network cards. Really. Not for any hardware advantage, but the drivers are solid, and there are serious flaws in at least a fair number of Linux network drivers. In both the tulip and rtl8139 NIC drivers, I've found that the driver *fully* resets the card when a transmit fails (produces a collision) 16 times in a row. Basically, that means that on nonswitched network segments under *extremely* heavy load, you'll get lots of "transmit timed out" messages, and then you'll lose network connectivity for a second or so for each ocurrence. I read up on this, and these are not unique to me. This can make Linux totally unusable on the network. 3c59x doesn't exhibit this problem.
Also, if *I* submit a proposal for IBM to buy me my dream computer and mention "Linux" somewhere in there, does that mean that they'll buy *me* one from that 1 billion dollars allocated for open source too?
3DLabs is pushing Unix technology by helping advance openGL. Check dri.sourceforge.net and you'll see the GMX 2000 supported. On eBay, you can get a used, but mint AGP 32MB GMX2000 for around $50. That's 100 fps sustained, which one of nVidia's or ATI's gaming videocards would croak under the geometry load of what's being rendered.
ATI and nVidia are real nice, but lets get some more heterogeneous Linux systems, not a bunch of mono-toned, spit-dribbling babies crying about how good their performance on their nVidia GePhorse versus them not needing open-sourced drivers. Linux was meant to be 100% open-source.
Everyone with an educated brain can see the bugs and give a crack at them if they may: Linux's key to progressive stability. But hey, who said consumers had to be smart to say "nVidia rocks, I don't care about open-source drivers. I play games. Look at the cows and eggs. Aiieeeeeee!"
But I'm sure you already Gnu that.
They failed to mention the Avant Stellar, the best keyboard out there expensive as keyboards go but this this has gone through several coffee and coke spills, a dunk in warm water is enough to bring it back to life, it also has function keys down the left hand side and is fully programable. One mean keyboard.
My current setup consists of three rack mount PCs, one rack mount SCSI box, three tower case PCs, two desks (monitor, keyboard, mouse on each), network laser printer, ADSL and cable modem, 100BaseT switched LAN, DDS-4 and DDS-2 tape drives. Total cost perhaps around 20-30k USD at prices when the stuff was bought. Oh yes, two laptops, but those I don't own - they are for work.
I run Linux and Windows on these computers. Because that's what I need for my software needs.
I don't see why I would run Solaris on PC equipment - if I need a Solaris box, I should probably get some cheap UE220R or UE250 for that (try dotcom sales for cheap sun hardware). Same with other OS'es - if I want a MacOS, I'll get a Mac. If I want OpenVMS, I'll get an Alpha. And I don't want HP/UX, so I'm not going to get one.
What's most expensive? Storage subsystems. Diskspace is just plain expensive. Even with IDE disks, a rackmount 4U computer case with 7 IDE drives in removable bays and an IDE RAID controller cost about 2700 USD. And that's the absolute cheapest way to get about 240 visible, redundant, fast, reliable gigabytes of storage.
Next most expensive thing? Networking. A 100BaseT switched LAN just doesn't cut it - it's gotten slow. Of course I could just get enough diskspace for ALL computers, but that's expensive, too, so I use a lot of network storage. Which puts real strain on the network.. Let alone trying to do anything serious over the network..
Of course everyone has different needs. Don't ever even think of AV work as a hobby.. Digital video and audio equipment (the pro-grade) costs an arm and a leg, and make-do equipment has serious performance bottlenecks.
All in all - a decent new computer would cost me about 10-20k USD. However, if I'd just want to play the latest, gratest 3D FPS games, the dream setup would be a lot cheaper, coming to perhaps 3-4k USD. And that's what most people consider the "expensive computer needs" category. However, that's because I already have about 15k USD in my AV rack. "Dream" gaming station should, IMHO, include good quality audio hardware, which probably costs much more than the computer.
Now, I know that I may not be a "Real Linux User", having used Linux only from kernel 0.98[some letter], only using it for work and hobbies, not having written more than three kernel drivers (subcontracting, and for custom hardware that most of You have ever seen or will ever see). But, for me, the ULB is much more than for ESR, as I care about my storage subsystem's reliability and speed more than most. It's usually the worst bottleneck, and there's never enough of it.
And yes, two of my Linux boxen are indeed old junk that wouldn't run newest WinME/2k/XP with any speed that we could talk about. They're external network connection gateways and thus don't need to be fast.
You're thinking of inexpensive ATA RAID, while they explicitly wanted a SCSI solution for speed. But SCSI RAID is _expensive_ - it's professional workstation class hardware, not within the budget for a personal machine (no matter that they say "cost is no object" - clearly there are limits here).
You either do RAID 0 Mirroring or RAID 1 Striping (with 2 ATA drives) or RAID 0+1 Mirroring and Striping (this takes 4 drives). Full striping plus Parity-Checking is RAID 5 but (someone correct me if I'm wrong) this isn't available for inexpensive ATA disk arrays. It would be nice if it were, but it would be slower than using a couple of SCSI disks and taking regular backup images of them. (What's best for backup is for yet another discussion.)
RAID 5 can be had for SCSI disks, at impressive prices, at which point you're better off with Gb Ethernet or Fibre Channel NAS or SAN storage. To do RAID 5 right, you need (some multiple of) at least 9 disks (8 for data, 1 for parity, with data and parity stripes randomly assigned across the array). The RAID 5 stuff gets rather complicated and expensive (have you priced SAN storage lately? I have, and it runs to 5 or 6 figures to just get started).
I like their approach for a high-end Linux machine for personal use. I'm using something similar as I write this (Tekram SCSI adapter with two 10K RPM Quantum 9GB non-mirrored disks). They're right to focus on I/O speed as more important than CPU power. Net bandwidth is the real limiter.
In this, they're just following what was learned long ago on mainframes: tune the I/O subsystem first because that's where you find large delays, then make sure you have enough memory (since Virtual Storage impacts Real/Expanded Storage, which impacts Auxiliary Storage - back to I/O), then tune CPU allocation and capacity last. It's well known that when you finally run out of CPU power (having tuned in this order) it's time for short-term triage (favoring "loved ones" at the expense of discretionary workloads) followed by an inevitable configuration upgrade. This is how it's done, folks.
But! You can now buy off-the-shelf parts (here, for example) that all work together and can just be bolted together. You can build sealed systems, removing the risk of spills if you move the machine and meaning you don't have to top the system up to allow for evaporating levels. You can get dinky little 120mm radiators which can be fitted inside the case, meaning the entire system can be self-contained. And if the system is well-built enough, the risk of a joint bursting and soaking your motherboard is a lot less than your HSF falling off and frying your Athlon.
Balanced against that, you can get cooling performance superior to a fan-based system and a hell of a lot quieter. And the disadvantages of watercooling will only get less as they become more and more commoditized.
You win again, gravity!
I've been having problems with the Thunder K7 hanging under heavy loads when running with 2 processors. Similar to Eric's problems, but I don't have a sound card to blame it on.
It doesn't seem to hang under casual use, but hangs only under heavy load (long make -j2 compiles). I suspected hardware troubles, but everything is running cool and I've swapped all pieces with the same results (NMB power supply, SIMMs, CPUs, motherboard)
Has anyone else had this problem, or found a solution?
jeff
The 1.7GHz Xeon processors that you use at work are not P4's though, although to be fair, the option to use Xeons should have been mentioned in the initial article. It is perhaps important to note though that the current Xeon (such as the one that you have at work) is essentially identical to the current P4 except for the different socket used and the fact that it has been validated for SMP use. The actual processor core is the same.
That being said, the price tags of the P4 and the Xeon shows a definite difference between the two chips. Combine that with the fact that Xeon boards are quite expensive, and for the time being they require fairly expensive RDRAM memory, and you get a rather costly setup. However, in situations where cost doesn't matter as much as performance (which is what the original article was trying to get at), the dual Xeon would definitely be the closest competitor to the AthlonMP.
Now, as to just how close of a competitor the Xeon is to the AthlonMP, well, if you read this article at Ace's Hardware (one of the very few hardware websites run by people who actually know what the heck they're talking about), you'll notice that dual 1.2GHz AthlonMP's match or beat the dual 1.7GHz Xeons pretty much across the board. The high-end workstation market is one area where AMD currently has Intel beat solidly in terms of raw performance, and pretty much trounces them in price/performance. Of course, that hasn't stopped Intel from maintaining 95%+ of the high-end workstation marketshare... but I digress.