nfs server tuning, bdflush, 2.6 kernel

[depontius]

I run /home from an nfs server, and have been suffering from firefox/sqlite for some time. I've seen an article about nfs server tuning, and they suggest tuning /proc/sys/vm/bdflush to make the server do more frequent writebacks, and reduce the likelihood of nfs activity colliding with the every-30-second writeback, causing a pretty bad storm. I know that the real problem/solution is with sqlite, policy, etc, but the bdflush certainly looked like a good stopgap/mitigator.

I have no /proc/sys/vm/bdflush - this looks like it went away somewhere during the 2.4->2.6 or 2.6 itself. I seem to pick up that there are 2 dirty age parameters that pick up some of the function, but haven't found anything about how to translate the old /proc/sys/vm/bdflush parameters to a newer kernel.

Can anyone suggest or point?
_________________
.sigs waste space and bandwidth

[slackline]

Some similar problems here, not sure how useful it will be though.
_________________
"Science is what we understand well enough to explain to a computer.  Art is everything else we do." - Donald Knuth

[depontius]

Not very, but thanks for trying.

The problems with firefox/sqlite on ext3 are pretty well-known, and if HOME is on nfs, the situation is considered just about paranoid. I should probably be considering one of the methods of moving "~/.mozilla/firefox" either back onto local disk or onto tmpfs, but I'm not really happy with any of the methods I've seen. I've also found that when firefox/sqlite is creating an nfs storm, one way to lessen the impact is to force some other nfs traffic in an xterm, with something as simple as "ls -l". This gives me the gut feeling that a little tuning won't fix the problem, but will lessen it enough to make it livable until firefox/sqlite finds a cleaner way to make sure their bookmarks are safe.

On looking around I found directions on kernel tuning on nfs servers, and they advised changing bdflush parameters. But for Linux 2.6, bdflush has been replaced with pdflush, and of course the parameters are all different. I've since figured some of this out, and begun searching for documentation on tuning pdflush. There isn't any easy "do this for nfs servers" like there was for bdflush, so I've got some reading to do. But in the meantime, "powertop" always suggests bumping dirty_writeback_centisecs from its default 500 up to 1500, which is the right thing to do for a laptop, and I guess the right thing to do for power consumption, but the I suspect is wrong thing to do for a server. I've undone that change and printed the documentation for later reading. In the meantime:
http://www.westnet.com/~gsmith/content/linux-pdflush.htm
http://www.dba-oracle.com/t_tuning_linux_kernel_2_6_oracle.htm

This isn't solved, and I'm still eager to hear any suggestions anyone has to give. But as I figure things out, I will put it here for others to see.
_________________
.sigs waste space and bandwidth