I agree with your reasoning - cheap motherboards are readily available.
For use at home, however, I looked at the mini-ITX motherboards. Less CPU performance of course, but low power and low cost.
I don't think there'll be power problems with such a small number (initially just 2) of slaves, really. I'm constrained by a fairly tight capital budget, designed to make 1 good PC, and I found I could turn it into one respectable PC plus two cheap slaves.
If I really wanted to push the CPU-power envelope, I'd get a refurbished PowerMac. I still might. The cluster would give me an aggregate 2500 MFLOPS, while a previous-model Mac would give me 3500 for about the same price (both figures being what the compiler alone can do). The only thing holding me back from that, is the possibility of needing to run high-end physical-simulation software that is probably only available for Windows. (Yes, I know about PC emulation software, but it takes a large performance hit.)
BTW, this setup is officially to allow me to move my personal machines off the desk and back home! They're currently an Athlon 1.2GHz and a Duron 800, plus a Pentium-MMX and a 486 that are used as servers.
I also have a PC-chips 810MLR, which we used for Linuxbios work last summer. I do hate to say it, but Linuxbios does work, but PCChips have a rather bad reputation.
Given the price, this isn't *so* surprising, but given the price it's also easy to "forgive and replace". Because they are slaves, it's easy to take one offline if it goes bad, and take time over getting it replaced (probably under warranty). I hope LinuxBIOS is more flexible than the rather anaemic status page suggests.
If someone else is paying, from a grant, then think about getting a cluster with better hardware, say using Tyan or Supermicro motherboards.
I will have to see if any are available at a comparable price. I do know that finding a case big enough to fit a Tiger into, while still keeping the price down, could be a real pain. But then, SMP boards tend not to be designed to be cheap. :)
That is a good point though - I should look at putting a pair of SMP machines together, and see how the price compares. However, I think the single SMP I priced was scraping the budget already - Athlon-MPs are priced noticeably higher than their single equivalents, which definitely hurts when compared to clusters of single-CPU slaves.
A little more capital cost will result in much less grief later. And also think about the cases - do you have room for a pile of big ATX cases?
Let me see, right now the desk has one full-tower and one midi-tower ATX, one AT, and one low-profile desktop, plus two monitors, at least three keyboards, plus a laptop, all piled on and around it. I'm planning to replace all that with one large medium-tower and two midi ATXs, plus one monitor and one keyboard. I don't think space is a problem. :-)
Where is your department?
Lancaster, UK.
On Mon, 24 Feb 2003, Jonathan Morton wrote:
refurbished PowerMac. I still might. The cluster would give me an aggregate 2500 MFLOPS, while a previous-model Mac would give me 3500
how did you get to that #? just wondering. If you're counting on Altivec that's a mistake.
ron
refurbished PowerMac. I still might. The cluster would give me an aggregate 2500 MFLOPS, while a previous-model Mac would give me 3500
how did you get to that #? just wondering. If you're counting on Altivec that's a mistake.
I count Altivec because I'm confident the Apple-modified compiler is capable of producing code for my present algorithm. Even if it can't, I can write it in fairly easily - I'm no stranger to PPC code.
On Tue, 25 Feb 2003, Jonathan Morton wrote:
how did you get to that #? just wondering. If you're counting on Altivec that's a mistake.
I count Altivec because I'm confident the Apple-modified compiler is capable of producing code for my present algorithm. Even if it can't, I can write it in fairly easily - I'm no stranger to PPC code.
our experience has been 'don't believe it till you try it', but maybe it will work for you.
ron
-
Jonathan Morton -
Ronald G. Minnich