Moreri has been doing this super-high-load thing for a while. Having nothing to lose, I've upgraded it to 2.6.
Before: up 18 days, 23:42, 1 user, load average: 48.34, 74.74, 100.16
After: up 25 min, 2 users, load average: 1.08, 1.36, 0.86
:)
It's running 2.6.6-1.376smp, from rawhide packages. I don't forsee any real problems, so will probably upgrade the rest of the web boxes like we have the others soon.
-- brion vibber (brion @ pobox.com)
On Sun, 23 May 2004, Brion Vibber wrote:
It's running 2.6.6-1.376smp, from rawhide packages.
How many times do I have to tell you people to build your own d*** kernel and stop using the junk from the distribution (esp. redhat.)???
(On y'all's :-) hardware, it should take less than 10min to build everything you actually need. It takes me 25min and I build everything.)
--Ricky
Ricky Beam wrote:
On Sun, 23 May 2004, Brion Vibber wrote:
It's running 2.6.6-1.376smp, from rawhide packages.
How many times do I have to tell you people to build your own d*** kernel and stop using the junk from the distribution (esp. redhat.)???
I don't suppose you have any actual useful information to give about the actual set of kernel config options that will get DMA working on our machines, do you? Believe it or not, *it doesn't matter* who compiled the kernel if it doesn't have the right driver in it.
If all you've got is "Red Hat sux0rz", please do us a favor and don't post.
-- brion vibber (brion @ pobox.com)
Brion Vibber wrote:
I don't suppose you have any actual useful information to give about the actual set of kernel config options that will get DMA working on our machines, do you? Believe it or not, *it doesn't matter* who compiled the kernel if it doesn't have the right driver in it.
He has a generally valid point that slimmer, non-generic kernels are a good idea on production systems. As for DMA, has anyone tried recompiling a few times and seeing if they can hit the right options? Is there a non-production machine where this experimentation can be done without danger?
I run a (1+5)x100GB RAID5 storage server, and when first installing it, couldn't get DMA to work for the life of me. I ended up writing a quick kernel hack to make it work on my system, and I think the situation was fixed a few kernel releases later. If you need me to look at the machines and see what I can do, let me know.
Cheers, Ivan.
Ivan Krstic wrote:
He has a generally valid point that slimmer, non-generic kernels are a good idea on production systems.
It's not a bad idea, but waving it around like it will cure cancer is pretty useless.
As for DMA, has anyone tried recompiling a few times and seeing if they can hit the right options? Is there a non-production machine where this experimentation can be done without danger?
Yes, I've tried recompiling 2.6.6 on curly without success so far (that was a week or so ago).
I wrote to Silicon Mechanics to see if they've got any info, and they claim that:
If you upgrade the kernel to at least 2.4.20-28.9 (this is what I tested with) DMA will be enabled by default with the default BIOS controller settings. I highly recommend doing this anyway because there are a number of local root exploits in all but a few of the very latest
kernels.
Unfortunately that doesn't help much since it *doesn't* work on a shiny new 2.6.6 kernel so far as I've been able to determine (either custom compiled or from Red Hat). Of course if the BIOS settings *aren't* the default, how are we to know? I asked them what might need to be changed, I'll ask for more details...
-- brion vibber (brion @ pobox.com)
wikitech-l@lists.wikimedia.org