Slashdot Mirror
How Many Desktop PCs Can One Server Replace?
NZheretic asks: "HP has just announced that they have upgraded a four-processor server with Advanced Micro Devices' new dual-core Opteron. The amount of processing power a multi-processor multi-core system can deliver seem like a waste of processing power for most traditional servers, which are more likely to suffer from disk access bottlenecks before lack of processing power becomes a problem. But what if that power could be delivered direct to the desktop users? The HP ProLiant DL585 supports eight 64-bit PCI-X I/O Slots (Six 100MHz, two 133MHz). The ATI FireMV(TM) 2400 supports Quad DVI/VGA displays on PCI Express. Assuming that you leave one PCI-X slot for a multiport USB card, thats up to twenty eight displays with USB keyboards,mice and headsets that could theoretically replace twenty eight networked desktop PCs. Using DVI and USB extenders, not all of the user stations would have to be within the 7.5 meter cable distance imposed by the DVI cable limit. The only OS currently capable of supporting this many displays is Linux. What limits would be imposed by the hardware and PCI-X bottlenecks? Taking into account the added cost of the HP and ATI hardware, could it deliver a great reduction in the total cost of ownership over both traditional PCs and thin client systems? How many desktops is it practical for a high end server to directly replace?"
I'm sorry, but this is one of the dumber products I've seen out there.
The software retails for $99 per workstation, and this gets you only one year...additional years are $29, again per station.
Add to that cost the cost of all those dual-headed video cards, USB cards and hubs, and DVI and USB extenders, and your total cost is not at all inconsequential. And for all this work, you have a maximum of 10 users to a server? Plus, those users are physically tethered to the server, severely restricting your network design.
It seems to me that all this and a lot more could be accomplished with less money and less hassle via some very low-end systems and VNC. In fact, that's how I'm accomplishing it right now.
____
~ |rip/\/\aster /\/\onkey
..and on my desk, then I'd say 1.
-m
http://www.invisik.com
Can't you already set up Xorg to configure which monitors are with which keyboards/mouse etc..
The only issue would be the USB/user thing... but with hald/gnome-volume-manager (not sure of the KDE equiv), this can be worked around...
Why buy this product when I _THINK_ you can do this already.
The power can be direct to users via Linux Terminal Server Project. Use a gigabit network and you can have lots of users. But why would someone buy it if it has too much processing power for their needs?
Haven't we gone through multiple iterations of this idea already? Dumb terminals, thin clients and now well...dumb terminals again. I mean you could do it but isn't this just a rehash of a really old idea?
Pci-Express and PCI-X are not interchangable. PCI-X is really fast PCI, where-as PCI-Express is different altogether (Although a PCIe to PCI/PCI-X bridge is supported).
Depending on how these systems are configured, it may not be possible to use that many monitors.
...a beowolf cluster of these?
PCI Express is normally shortend to PCI-E or PCIe
Your plan will not work with this motherboard.
Linux treats all keyboards and mice as a single input source, so you'll need to get patching if you want more than one active user at a time...
Take that giant server and put it in a back room under lock and key. The only things that should plug into it are a single power cable and network. Put a single KVM in your rack to access all the servers in it.
Now buy 30 thin clients. Each one gets a KVM and a network card. Good. Now plug in the power on all the thin clients and plug their network cables into a switch. To remove clutter if you want you can use 802.11 and all the thin clients will only need power.
Ta-da! Welcome to intelligence.
The GeekNights podcast is going strong. Listen!
OK. Given that a two processor version of the DL 585 is $16,000US and does not include any storage, we can assume that a fully loaded box processors, memory and some storage, is going to run $28,000 plus and that doesn't include monitors. That's more than $1,000 per user just for hardware. Since the average business PC runs under $1,000 the server solution that you suggest just isn't cost effective.
:(){ :&:;};: at a bash prompt and you're fired.
Now add to that cost, the single point of failure issue. Even if the hardware never fails, all you need is for some malicious or clueless user to run
The cost sounds pretty reasonable for a corporate installation, and I think it's safe to assume that with a subscription, you get the software updates and patches as part of the deal. For the type of business that will buy this thing, these numbers just are not that bad. HP products are not "free as in beer".
"Who are in control, they are not in control of anything - they don't even control themselves!" - Glen Beck
Why do you need an USB card? The server already comes with 2 USB ports, and an USB bus supports up to 127 devices.
1. Click on the word "preferences" in the first column, right below where your slashdot user id appears.
2. Click on the "Homepage" sub-tab
3. On the right, in the second section of the homepage preferences, uncheck the box next to the word "Cliff"
4. All done!
I forgot to metion that the same super server could service hundreds of concurrent users if you used Linux or Citrix with thin clients. Depending on your applications disk io is not the big bottleneck. Usually when it comes to terminal server bottlenecks it is memory, processors, and then disk io.
When you do it this way, the cost goes down in dramatic fashion. A $50,000 server setup is only $250 per user when you have 200 users running off it and A server as large as you suggest could easily run 500 or more users concurrently, which would further bring the cost down to $100 per user!
Wow, lots of factual mistakes. PCI-X != PCI Express. PCI is a bus, PCI-X is a faster version of that bus. PCI-Express is next-gen AGP. I don't know of many PCI-X video cards. As for input, seperate devices are marked seprately in the kernel, if you just use /dev/input it conglomerates all the inputs. You still need a decent Xserver/servers to handle all the seperate sessions, though.
And since X is monolithic, you'll need to run seperate X threads per display, or one idiot going to a website with a thousand animated .gifs will stop everyone.
In short: Bad Idea.
How many desktops is it practical for a high end server to directly replace?
None, just like a big truck doesn't replace any passanger cars.
You could, however try something like OpenVPS to replace a couple 'o dosen servers with it...
No one makes PCI-X display adapters, only regular PCI ones, and those are getting harder to find and unsuitable for what you desire.
This machine has 0 AGP and 0 PCI Express slots, only "Graphics Integrated 1280 x 1024, 16M color on PCI local bus, 8 MB of SDRAM video memory".
Neat idea, but sorry.
PCI-X is not PCI-Express, and the two technologies don't even have compatible pinouts or signals. PCI-X is the follow-on to traditional parallel PCI, with speeds of 100 and 133 Mhz and a 64-bit wide data path (compared to previous parallel PCI standards of 32/64-bits at 33/66 Mhz). PCI-Express is PCI re-done serially instead of in parallel, in an attepmt to be fashionable like the new Serial ATA standard. It's also potentially faster than PCI-X, and not at all compatible.
You'll notice just about every communications standard that doesn't go long haul alternates back and forth between parallel and serial methods every few years just to sound new and exciting and better.
11*43+456^2
Actually, one quad opteron server can be a terminal server to about 600 clients...:) Doing it now! Saves about 40hours a week in "spyware" issues..
TIME is the Aether...
When you have a separate system for each user, you're pretty much relegated to using NFS/SMB for network data storage. But with a single system for a couple of dozen users, it suddenly becomes much more practical to put a fibre channel card in the system with a direct connection to a storage system. So the assumption that I/O would become a bottleneck in such a system may actually not only be wrong, it might be faster.
What about using VMWare to give each user their own system? Then not only can you run different OSes for each user, including ones that don't support multiple keyboards, but you also can let users move virtual machines between physical work areas (or even between different servers).
I know that the higher-end VMWare products support migrating between servers. I don't know how well they support multiple physical input and display devices, but I suspect that if a major customer requested it, it could happen.
Disclaimer: I work for VMWare's parent company, but I have no inside information on the product, as I'm in a totally different division.
Mega-systems like this make sense for big SQL and application servers where CPU is a bottleneck. Most of the other uses are solutions looking for a problem. Intel and AMD would dearly love to corral more of the dollars spent on building systems; these mega-chips have high margins. But the reality is that it is going to be more cost-effective for most setups to use a large number of inexpensive PCs with modest CPUs.
How the hell is this any different/better than using SunFire servers and Sunray thin-clients? A Sunfire V480, 4 900mhz UltraSParc III processors, 16 GB RAM, mirror 73gb disks. This system ran 100+ Sunray thin clients all running continuously updating graphical simulation displays with 3 or 4 other semi-rigourous processes, on a 100mb LAN. All data and programs were NFS mounted. The V480 was ~$40-50K + $300/sunray (already owned monitor and file server).
The system was spec'd by Sun to handle those 100 sessions. The head engineer bought two and set them up to load-balance and provide redundancy.
This isn't anything new...move along.
"Get a bicycle. You will not regret it, if you live." - Mark Twain, "Taming the Bicycle"
There's a product called Buddy that's been doing this for many years. Originally, the Buddy card was a combined PCI video card and PS/2 keyboard+mouse controller, which spit all the signals out an 8-position modular jack (RJ45 for the cretins), and a little breakout box at the other end of a (long, shielded cat-5) cable accepted the monitor and input devices. The software gave two Windows95 users the impression that they were the only one on the machine, and I'm still not sure how they did that on a non-NT architecture, but it worked and worked well. Only trouble is, the video bandwidth of the cable was limited, and the RAMDAC in the video card didn't support sync rates over 60Hz, so the flicker on the slave station was pretty obnoxious.
In the years since Buddy was first released, PCI video cards have learned to play nice with their neighbors, and USB has provided a way to connect oodles of keyboards and mice to the same machine. Thus, Buddy is reincarnated as BeTwin, a software-only product that associates specific keyboards, mice, and video cards with specific sessions on the machine. (I'm not sure how it deals with sound. Multiple soundcards would seem easy.)
They say it only supports 5 users, but that sounds like an arbitrary limit and I'm sure they'd tweak a 28-user version if you felt like it.
Related links... I'm going off-topic here, but playing stupid tricks with virtual hardware is fun.
Check out MaxiVista, a "virtual video card" which Windows treats as a second monitor, allowing you to do multi-head tricks. The data for the second display goes out over the network (a la VNC) to a client machine, which simply pipes it into the video buffer. Turn that scrap laptop into a second monitor! I stuffed a 10base-T card in my old lappy and it was perfectly usable for everything except fullscreen video. At 100 or gigabit, it'd be worth a try.
Xinerama is Linux software that does the same thing, creating one large virtual X display, which then chops up the image and sends it to a number of smaller actual displays, some or all of which can obviously be located across the network.
As long as we're doing silly tricks with virtual hardware, you should be aware of Virtual Audio Cable, which enables digitally-perfect audio patching between applications' outputs and inputs, even if the apps themselves think they have exclusive control over the soundcard. (Also enables multiclient sound output under 9x, even if your card doesn't support it, because it does software mixing.)
If video is your thing, try Softcam, to feed your videoconferencing software any old source you feel like. Switch between actual cameras, use your desktop screenshot as a "camera input", add effects, etc. Their WaveMux tool is a nice complement to VAC, too.
HOWTO: Multiple local XFree users under Linux
Google Link (Linux planet if you want to jump to the particular howto)
Idea is X directly setting which display to use, and setting which usb device to associate with.
e a lot of time, but seems easy enough for 2 people on a box.
(This is more to all the nay-sayers that seem to imply several people on one box isn't doable)
Huh? It depends on the application. You could run 1000 dumb terminals in a call center with a single 900 MHz P3 based machine. Old Sun SPARC 2 can handle 64 dumb terminals without breaking a sweat.
Perfect correction, parent!
In the old days, there was one big (relatively) powerful computer with a bunch of terminals hanging off of it. This computer was called a Mainframe.
As time went on and miniaturization progressed, people wanted their own department computer, so they could have more CPU time available.
Then they wanted their own desktops.
Then they wanted to network their desktop machines, so they could share data.
Then some applications started sharing CPU and other resources over the network.
But all these networked machines were a big configuration hassle for IT. They envisioned "thin clients" and similar solutions.
Now machines are so powerful that users can have their own virtual PCs running on a central server, so they can just have dumb terminals on their desks.
There's a lesson in here somewhere. As soon as the network comes back, I'll google for it and find out what that lesson is.
Eloi, Eloi, lema sabachtani?
www.fogbound.net
It includes keyboard and mouse and NO moving parts.
Eh? Last time I checked, clicky buttons are considered moving parts. So either you're exaggerating or you're saying you got some sort of electrostatic keyboard and trackpad?
I recall there being a post on /. a few months ago about how this has already done to a lesser extent -- on Linux -- in (South?) Africa, where one PC is providing displays, keyboards and mice for four simultaneous users to cut down the cost per seat. I'm sure the information on how they pulled it off is readily available online with a little bit of digging.
I guess it depends on how narrowly you define "traditional servers". Maybe you meant file servers. But plenty of database servers, and most compute servers, are reasonable candiates for this.
And there are plenty of those around.
We use lots of compute servers. After reliabaility, we care most about TCO per CPU/RAM set. A dual Opteron with 16GB is cheaper than two Opterons with 8GB each. But even if it was slightly more to buy, we'd take two in 1U over 2 in 2U. This scales forever, or until we hit some practical limit.
We've always run our db servers on dual processor systems; I suspect we'll welcome the quads at some point in the next year or two.
Why not have robust distributed operating systems?
Each X11 server (which is what people would be running) can be configured to use whatever collection of keyboards, mice, and displays you want it to use.
Of course, that makes it no less of a stupid idea to do that (you should be using an X terminal and set the thing up as a server). But, in principle, Linux will support this sort of insanity if you must.
I've thought about this a lot. Now that we have SMP and SLI graphics cards, it's realistic to consider the following: if you want two good computers in a house, is it better to get two separate systems, each with a good processor and graphics card, or is it better to get one multi-user system, with two processors and two SLI graphics cards? I think the second is obviously better, and it doesn't have to cost more (it does now only because ordinary people are not buying these yet). It's better because even if both stations are in full use, each will still have the processing/rendering power of one isolated system. For the times when only one is it use, it will have a lot more power at its disposal. The same reasoning applies for a 4-way system, etc. I think this is the way of the future. The quietest computer is one that lives in the basement. I hope that future houses will be designed like that: A big "mainframe" which sends I/O to monitors, peripherals and amps, but also the TV and stereo in the living room. This big computer should be somewhat expandable, so you can add CPUs and graphics processors when it's time to upgrade. Yeah, that would be awesome.
Cluster the cheap desktop computers, add some expensive glue and storage that keeps things reliable and run stuff like OpenSSI.
Simplistically speaking you split the desktop computer into two. One is part of a "Big Server", the other is the "Thin Client".
If multicore CPUs and virtualization becomes common this isn't going to be that hard.
Of course if users randomly pull the plug on their nodes that does make things a bit problematic. So I suspect the current "thick desktop" stuff is going to be around for a while.
Given the thick desktop isn't that expensive, it's just _managing_ them that's usually the problem. One should just stick with the thick desktop, add those hardware cards that cause the computer to rollback to a known disk state on reboot. When hardware virtualization becomes common (pacifica, vanderpool etc) this will be even easier to do - you could be updating the client computer with a new image with updates etc whilst the user is using one image. No need to pull the whole computer off line while reimaging it. You just reimage one of the "virtual computer"s. The next reboot would run the updated virtual computer. Dual core CPUs etc will actually make this not too bad.
The persistent storage, 24/7 processing, and backups would be on the main server/cluster.
Matrox has a quad monitor PCI card based on the G200 chip. They even had an 8 monitor card based on the G100 at one point. Not sure of Windows or Linux would support six such cards though! :)
Here's a 4 port version on eBay:
Matrox G200 MMS G2+ QUAD PCI Video card/cable
As others pointed out, your plan won't work because PCI-X != PCIe.
However, Matrox has made 4 and 8 port PCI video cards for many years. Here's is one such card based on the G200 chipset (on par with the NVidia TNT2 performance).G200 MMS on eBay
I've been looking for a cost effective X terminal solution but so far, all the clients that I have found are $300 and UP. What x term did you find for $150 and where did you find it?
Never mind running multiple users off it, I want one for a dedicated flight simulator.
I'd been looking for that device a while ago, but I didn't find anything promising. It still seems like vaporware, I'm unable to find anyone actually selling it. For those whose VGA ports are on the docking station, the USB2VGA would be a better way of driving a projector from a small laptop. Or driving a whole pile of monitors at once.
Being USB 2.0, I'm surprised there's a bandwidth problem. I was running MaxiVista at 10Mbps, and it was tolerable. At USB1.1's 12Mbps, I'd expect similar. But at 480Mbps? It should be capable of full-screen movies. I'm pointing my finger at the USB controller or drivers.
Anyway, most of the USB video output devices I find are made to drive NTSC composite or S-video. That could also be fun for a lot of applications.
I used to build systems a long time ago that ran PC-MOS on 386 boxes to run multi-user MS-DOS systems. Combined with a (at the time) super sexy Maxpeed ( http://www.maxspeed.com/ ) it would allow you to run DOS programs on serial terminals like the Link MC5 and similar.
Pretty cool stuff for the time. Also awesome for running multi-node BBS systems.
--> Fight tyranny and repression.... read
Windows 2000 TSE and 600 users?
2 gig limit per process?
You are smelling a lot like a troll
Your Average Joe