Slashdot Mirror
IBM's $45 Linux Server (Well, Kinda)
Wedman sent us a snippet from a newsletter from
OpenSourceIT that starts off by saying that IBM will announce a new pricing scheme for Linux
on the S/390 mainframes: Soon they'll cost $125k. For another $20k you can get virtual machine software to run multiple copies of Linux on the same box.
David Boyes, a consultant who works with the S/390, managed
to boot 41,500 Linux servers on one mainframe. Although he notes that
you may not
be able to run that many in real life. ;) (if someone can find an actual link for this, please post it) That just cracks me up: I mean, the debate about forking apache to handle requests is one thing, but hell, why not just boot your own OS for each request!
That just cracks me up: I mean, the debate about forking apache to handle requests is one thing, but hell, why not just boot your own OS for each request!
I read the article on booting 40000+ linuxii on one box almost a year ago, so I might be a little sketchy, but if I remember correctly:
The issue is booting an independent copy of the OS for each 'instance' of a server, NOT FOR EACH REQUEST. This means that they could run, maybe 20,000 machines that look independent on one shared machine, each of the 20,000 virtual machines running independently of the others. 20,000 root accounts, 20,000 userbases, 20,000 sets of allocated memory, etc., all running simultaneously, off of the same machine.
Right now, web hosts can offer cheap web hosting (virtual hosting), where each user shares the OS, and with properly set permissions, and limited user functionality, this is relatively secure. This is generally run off of one, or a small group of IPs all pointing at the same machine, and the webserver figures out which 'instance' of itself should return what resource to the requester.
The problem is that, for instance, if I need to do something outside of my user-sandbox, I can't, or I need to have someone else do it.
This whole multiple instances of one OS one one bigass machine, appears to the user as a co-location. They don't have to worry about other users screwing with their stuff. Essentially, my instance of the OS on that machine is the same thing as my own box being hosted at the ISP.
AND, with virtual hosting, some user cracks root, and every account on that machine can be comprimised. With this, someone cracks root on one of the 20,000 instances, and whoever maintains that instance gets screwed, but the other 19,999 users are unaffected.
However, I suspect that we'll see people offering virtual hosting within an instance, which kind of defeats the purpose, but also allows, say 100 users to be hosted on an instance, which allows for 2million sites to be hosted on one machine.
I wonder if IBM needs beta testers (-: I'd re-wire my house if they sent me a demo unit.
I disagreee. PCs are cheap both in price and quality. The best-quality components are much better, but they still fail enough that with 100, 10000, or 41000 you will have several with problems at any one time.
Even with the best server-grade components, the PCs will be far less reliable. And I would bet that 10000 server-grade PCs would cost as much as one of these mainframes. If you own the mainframe, the cost-per-minute charges don't apply (remember that IBM sells now-a-days), and while you need more-expensive operators, you need fewer of them as they are not swapping hard drives and smoking video cards every 20 minutes.
For a long time I advocated clusters of PCs for any application. Clusters have some great uses, but so does real hardware.
One University (speak up if you know) is actually running thousands of seperate linux guests under VM on their s/390 and giving EVERY student their own Linux box to play with.
I'm under the assumption that they have some method of dealing with security in a central way so that everyone isn't running tftp and the dreaded r servers.
Finkployd
Don't fork the OS for each web request--fork one for each customer. Think about it--you buy ONE mainframe and ONE copy of Linux. You can 41,000 customers each with their own "machine". They can do whatever they want with it, including configuring the security themselves. It doesn't matter if they do it wrong, the other 40,999 customers aren't affected (with the possible exception of bandwidth).
--
Linux MAPI Server!
http://www.openone.com/software/MailOne/
(Exchange Migration HOWTO coming soon)
You could even implement a good response system to security break-ins. Any time someone logs in as or su's to root, indicate there is an error and swap to one of the hot spares. So what if the cracker trashes the one he is on. Switch to a backup.
Of course in this situation, you would need 100's of spares if someone is a little persistant.
Go not unto/. for advice, for you will be told both yea and nay (but have nothing to do with the question)
David Boyes, a consultant who works with the S/390, managed to boot 41,500 Linux servers on one mainframe. Although he notes that you may not be able to run that many in real life. ;) (if someone can find an actual link for this, please post it)
The story on NetworkWorldFusion News
The story on Fairfax IT
A reprint of the story from LinuxPlanet