"There has been a lot of discussion lately, both on
lkml and elsewhere, regarding the VM subsystem used in the 2.4
series of the Linux kernel. Recently, these issues have been
dramatically inflated by Linus replacing the entire VM (originally
built by Rik van Riel) with a whole new VM system by Andrea
Arcangeli in the midst of a "stable" kernel series, specifically
with the 2.4.10 kernel release.
Linus believes the new VM is more stable and "correct" than the
old one. Others, including Alan Cox, believe that Rik's VM is a
good direction and changing directions in the midst of a stable
kernel series is a bad idea. Still others just want to know "which
one is best."
This paper is not intended to answer any questions regarding
whether or not it was a good idea to replace the VM in a stable
kernel. Rather, it is an analysis done by a somewhat atypical
"power user" who uses Linux personally and at work to do all manner
of computing tasks (i.e. everything) and who wanted to understand
first-hand typical performance issues relating to both of the 2.4
series VM systems in comparison with each other and in comparison
with the previous "stable" VM system from the 2.2 kernel
series.
This is also not indended to deal with issues regarding the 2.4
VM systems prior to what are "current" versions at the time of this
writing. Several issues were floating around for some time in early
2.4 kernels that have since been fixed. People running distros that
use one of those original 2.4 kernels and wondering if the kernel
version may be related to any problems they might be having are
probably not going to see anything of benefit here. I didn't want
to do an analysis on seventy different versions of the kernel, just
the most recent."