Slashdot Mirror
Fitting A Linux Box On A PCI Card
An Anonymous Coward writes: "Running on Newsforge/Linux.com is a hardware review where Slashdot's Krow took a couple of OmniCluster's Slotservers and and built a cluster configuation inside of a singe host computer (and even had DB2 running on one of the card's inside of the host). Could something like this be the future of computing where for additional processing power you just kept adding additional computers inside of a host?"
Follow me here:
A computer used to take up a room.
Then, computers were large cabinets in a computer room.
Now, they are boxes in a computer cabinet in a computer room.
So we can extrapolate the next step for computers is to be cardss in computer box in a computer cabinet in a computer room.
It's a natural (obvious) progression really.
"Karma can only be portioned out by the cosmos." -- Homer Simpson
I am actually typing this comment on a Sun Microsystems SUNPCI card.. It's a celeron, I beleive a 466mhz or so, w/ 128m of ram. It has onboard video if you want to use an external monitor or can use the sun's native card if you want to run it windowed, ditto w/ ethernet. I've been using the card for about 3 months now, and other than some instability w/ SunOS 2.6 (which dissapeared in 2.8), I haven't had problems with it.. you can copy/paste between the Sun window and the 'PC' window, which is very helpful.. and though we are running WIN2000 on it (ok.. so shoot me) I don't see any reason why you couldn't run linux on it if you really wanted too.. All-in-all, the card is pretty badass..
//Phizzy
"Most European technology just isn't worth our stealing," -- Former CIA chief James Woolsey, referring to Echelon
don't see these things taking off for most uses because the PCI bus is limited to a measly 133 MB/S. Even newer 64 bit PCI slots found in some servers have insignifigant bandwidth to keep the data flowing fast enough to make full use of these things.
You've heard of Beowolf clusters, right?
Let's imagine I'm running some large routine to model some physical phenomena.. Depending on the problem, it is often possible to split the computational domain into small chunks and then pass only the elements along the interfaces between nodes.. So, how does that impact this discussion? Well, let's assume I can break up an NxM grid onto four subdomains. The communication from each node will consist of N+M elements (not NxM).. Now, let's take a look at our options. I can either purchase 4 machines with gigabit (~1000Mb/s) ethernet, Myranet (~200Mb/s) cards, or maybe I can use ip-over-firewire (~400Mb/s) to communicate between machines.. Gigabit ethernet has some latency problems that are answered by Myranet, but if we just look at the bandwidth issue, then ~1000Mb/s is roughly 125MB/s. That's slower than the 133MB/s you quoted above for a 32bit, 33MHz PCI bus.. Of course there are motherboards out there that support 64bit, 66MHz PCI cards (such as these from TotalImpact)..
You're right that the PCI bus is not as fast as the data io approaches use by IBM, Sun, SGI, etc to feed their processors. BUT, if I'm deciding between one machine sitting in the corner crunching numbers, or 4 machines sitting in the corner talking slowly to each other through an expensive gigabit ethernet switch, guess which system I'm going to look at?