> People should definitely take a look by the Stanford paper, but at the
> same time keep in mind that TCP numbers have gotten 2-3 times better
> since that paper, and NFS client read performance is up to 30 times
> better depending on what you do (if it's not cached, it's not going to
> be better, but...)
Great! Here are a couple more thoughts on the Stanford paper (I may just
be showing my ignorance here, if so, I apologize):
The paper stated that performance gains could be made (on the Pentium, at
least) in memory writes. I don't really understand this part of the paper,
and it sounds like this may only help performance on a particular
architecture. Could this be implemented, and is it worth doing? Would it
be a kernel or library feature?
Also, as I recall the disk cache has quite a bit of overhead, and ext2 is
significantly slower than ffs in manipulating large files. I realize that
the virtual memory and disk cache has been re-done recently, though, so I
may be out of date.
> Linux has gotten a lot better in the networking performance department
> especially in the 1.3.6x kernels (I've been working a lot on it, and so
> has Alan and company).
I'm always amazed by the kernel gurus. You people work miracles. Thanks!
-Dave
(Once again happily testing the newest kernels when not studying physics)