The HIVE Beowulf
My fried Cyrus and I decided to build a Beowulf out of spare and discarded computer parts resulting from many upgrades over time. The Beowulf, called the hive, is over at my friend Bill's place in Pioneer Square where we have access to a T3. The Beowulf is built entirely out of spare parts that were sitting either in my closet, or over at Cyrus's place. The only thing we did buy was the Intel Ethernet card in hive4 to try out the net-booting. All the other boxes boot off of a floppy. As the hive has grown, we've started receiving old computers from friends and other people to add to the collection. If you have some working motherboard, CPU and ram combinations (preferably with floppy + power supply to donate, email us!).
Beowulf's are super computers built out of commodity computer parts, some of the most powerful super computers in the world are Beowulfs. Ours probably won't hit any top supercomputer lists, but it will let us draw some computing cycles out of computers that would have otherwise been sitting in a closet taking up space.
We recently (8/15/00) made an attempt to get MOSIX running on the hive, which seems closer to where we want to be (one large virtual computer that just appears at a multi-processor system and can fork/exec onto the least loaded processor auto-magically). The initial reaction was fairly poor. We had some problems getting stuff compiled (the package seems a little too friendly), had problems booting the kernel (ie: after saying "Now booting the kernel" it didn't), we had problems with our libc on the hive (now glibc-2.1) and general frustration with mosix in general. This might be attributed to the fact that I brought down some beer to drink this time, but we'll find out for sure next time.
It's worth noting that our Beowulf will probably be out performed by a single pentium-III [UPDATE: This may not be the case anymore. we'll have to start trying to measure this soon].
The current hive is made up of 10 nodes.. it's not running PVM or any of the beowulf software that's built on top of rsh because that seems a little kludgey.. it is running beo-procps and bproc which looks like most of the future work we do will be based on since it seems a better long term route to take. The only thing the beowulf is currently processing is a distributed.net client working on cracking rc5.
The network has recently been upgraded to a 100M hub. I bought a 10/100 switch to install at the hive, but decided that I really wanted it at home instead, so the 10/100M hub from home ended up down at the hive. Another thing we'd like to do is start loading boxes up with open-bios so we don't need to use floppies. The one thing really holding this back is the very limited mother board support by open-bios.. to use it we'd have to use the SiS630 motherboard (which is a good one, has built in Ethernet I think) but we'd actually have to start buying hardware which seems contrary to the notion of using recycled computers.. I'm not sure I'll be able to stop myself from getting one or two though just to try it out.
[UPDATE: Cyrus has been looking around and only found very expensive intel motherboards that will support openbios. I found some SiS based motherboards (an Excel) for $99, but I'm not sure if that will work or not.]
Another thing I'd like to do is throw away the cases, and put together a DC
power plant to power the machines.. I've seen so many nice +- 48V DC power
plants in telephone companies facilities, I'd like to try doing the same
style power plant to supply the +- 5VDC and +- 12VDC the mother boards
need. Then we could just pop all the boxes onto some sort of rack mount
arrangement (sample design 
) and handle
higher density with less mess.
Here's some of my photos:
The next issue I think we might face is power. It's probably unreasonable to add more nodes until we can fit things into a smaller space and find something to power multiple computers..
5 Nodes, beoprocps installed, first go at getting everything running..
