Our SMP implementation scalability

[Dmitri Nikulin]

On 1/18/07, Kris Kennaway <kris at obsecurity.org> wrote:
I disagree with you about FreeBSD's SMP architecture being difficult
to develop and optimize (there are several really good tools for that,
like witness, lock profiling, pmc, and dtrace, etc), but at least
there are half a dozen or more people working on SMP scaling issues at
any given time, so there is steady progress being made.
The tools certainly help. But there are plenty of bugs left that are
just there because the locking is too complicated to follow, and
without those tools it's barely possible to predict the kernel's
emergent behavior. FreeBSD is a lot more reliable now than it was in
5.x days, but the extent to which the *limited* complex portions in
5.x confused developers for many months, combined with the fact that
the system is becoming more complex, not less, implies that it's only
the new tools that make the new complexity bearable.
I really like FreeBSD, don't get me wrong, but it's really hard for me
not to think that if the team had agreed with Matt's vision on SMP
then FreeBSD would be a lot simpler and more reliable, with the
manpower previously devoted to fixing complexity bugs being used to
dent Linux' dominance over many fields. I really really hope NetBSD
doesn't end up with similarly complex SMP, because that could really
kill it, with a minor fraction of FreeBSD's developer resources and an
onus to continue supporting many architectures.
---
Dmitri Nikulin
Centre for Synchrotron Science
Monash University
Victoria 3800, Australia
email: dnikulin at gmail.com