Linux Weekly News, Friday

Linux Expo coverage
Linux Weekly News
Contact us

Friday at Linux Expo
For whatever reason, Pacific HiTech does not have a booth here at Linux Expo. Not wanting to be left out entirely, however, the folks at PHT hired a group of people (described to me as "models") to wander around on the exhibit floor, attract attention, and hand out T-shirts. They got attention, all right; they also got booted out by show management. So they hung out by the doors of the convention center, until they got booted again. (The word we hear is that Linux Expo management threatened to shut down the IBM booth - with which Pacific HiTech was associated - if they did not desist. That would have been an interesting one to see them do).
So plan C went into effect: they staked out the convention hotels and passed out their T-shirts from there....

Ted Ts'o talked about "standalone device drivers." A couple of facts about Ted: his name is pronounced "Cho" as in "chosen." And he is about to leave MIT to join the folks at VA Linux Systems. Yet another top-tier kernel hacker has been snarfed up by the industry. Which is probably just fine; now Ted can be paid to work on Linux full time.
Ted's point in the talk is this: the size of the kernel has been doubling about every eighteen months. The biggest chunk of code in the kernel is all of the device drivers. Maybe it is time to stop tossing all of the drivers into the kernel directly; instead they should be maintained and distributed as separate units.
He held up PCMCIA as an example of how well a standalone approach can work. The PCMCIA drivers can track the (rapidly evolving) state of the hardware without being tied to a specific kernel version. You can get a new version of the PCMCIA drivers for your new card without having to run a development kernel. Some people actually don't want to run development kernels, for some reason...
The bulk of the talk was made up of details on how standalone drivers can be implemented; stuff about building modules, handling multiple kernel versions, etc. Useful stuff if that's what you're trying to do right now.

Jon "Maddog" Hall talked about real-world uses of Beowulf systems. It was a classic Maddog talk, wandering all over the place and full of good stories. Very hard to reproduce here; you had to be there, I guess.
Anyway, he had five reasons why Beowulfs are superior to regular "big iron" supercomputers:
5: They can be expanded more quickly 4: They are far less expensive 3: Maintenance is cheaper (supercomputers require an expensive contract from their first day; Beowulfs come with a three year warranty on parts). 2: Faster to fix (get replacement parts at Walmart) 1: You can actually get applications for Beowulfs
He gave a number of examples of uses for Beowulfs, ranging from ten-minute mammogram classification in Brazil to the simulation of asteroid collisions with the earth ("this sort of research is best done with simulation"). There are a lot of clusters out there, and, judging from the interest shown in the Extreme Linux track here, there will be a lot more of them in the future.
--jc