Kernel locking & the G5 Quad
Kernel locking & the G5 Quad
- Subject: Kernel locking & the G5 Quad
- From: Charles Gray <email@hidden>
- Date: Thu, 20 Oct 2005 17:31:51 +1000
Hi,
I noticed this morning that Apple have now released a quad processor G5
Power Mac. I'm curious if anyone has any details of how this interacts
with the funnel locking in the BSD part of darwin.
The BSDCon 2002 paper "Advanced Synchronization in Mac OS X: Extending
Unix to SMP and Real-Time" (Gerbarg) describes thw two BSD funnels for
files and networking. I also remember reading (though it would seem not
in this paper, and I right now can't find where) some other document
from apple which stated that adding more than two processors to a darwin
system had little benefit due to funnel contention and cross CPU
transfers for VFS and networking.
As far as I've seen funnels still exist in the latest kernels I've seen
at opendarwin.org. I'm just trying to sus out what's going on here,
since the apple web page lists definite performance benefits from the
quad CPU system. Eg...
* Has the VFS & networking slowly been getting finer grained locking &
been made re-entrant over the last few years so there is less contention
on kernel locking?
* Do apps now have less of a dependence on the BSD part of the kernel so
the funnels are no longer held as frequently?
* Do the modern CPUs have much better IPI performance so the overheads
of the lock contention aren't so bad?
* Or, quite simply, are the extra processors only of real benefit if you
have a very parallel task that won't hit VFS or the network (eg.
rendering, simulation, whatever), and the scheduler has been tweaked to
ensure a minimum level of performance (ie. not below that of a dual).
Anyway, just very confused :)
Chuck
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Darwin-kernel mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden