On Fri, 28 Dec 2007, Tiffany Snyder wrote:
What piqued my attention was the note that our forwarding performance
doesn't scale with multiple CPUs. Which means there's a lot of work to be
done :-) Have we taken a look at OpenSolaris' Surya
(http://www.opensolaris.org/os/community/networking/surya-design.pdf)
project? They allow multiple readers/single writer on the radix_node_head
(and not a mutex as we do) and we may be able to do the same to gain some
parallelism. There are other things in Surya that exploit multiple CPUs.
It's definitely worth a read. DragonFlyBSD seems to achieve parallelism by
classifying packet as flows and then redirecting the flows to different
CPUs. OpenSolaris also does something similar. We can definitely think along
those lines.
NOTE: 1) I said multiple instead of dual CPUs on purpose. 2) I mentioned
OpenSolaris and DragonFlyBSD as examples and to acknowledge the work they
are doing and to show that FreeBSD is far behind and is losing it's lustre
on continuing to be the networking platform of choice.
Thanks for your kind comments, we always appreciate it when people tell us
we've lost our lustre and are far behind.
A few more substantive comments--
Kip Macy has been floating a patch to move to reader-writer locking on the
routing code, but I'm not sure how mature it is. Certainly, we are aware that
routing lock contention is a key issue we need to look at for forwarding
performance. If he doesn't follow up to this thread, give him a ping for a
copy of the patch.
I have a prototype work dispatch system for TCP/IP on FreeBSD based on packet
flow classification, such as the ones you point at, which can be found in the
FreeBSD perforce server. However, life is a bit more complicated than just
identifying flows and assigning them to CPUs. In FreeBSD 7, we perform direct
dispatch of the network stack from the interrupt thread in order to avoid
context switch overhead, which mean there isn't a natural deferal point, such
as the netisr dispatch. In practice, for high performance devices, this is
both highly desirable in order to avoid dispatch overheads, but also because
10gbps cards almost always support in-hardware dispatch to multiple input
queues. This is much preferable pulling packet headers to a single CPU in one
interrupt thread, and then bouncing the cache lines for the packets to other
CPUs for actual processing. Our Chelsio 10gbps driver, for example, supports
work dispatch to multiple input threads today, so you will already see
parallel processing of input flows from that hardware via multiple ithreads.
For other devices, where either the hardware or device driver doesn't support
the in-hardware demux, having software processing is a good idea, but in order
for that to pay off, routing lock granularity and TCP input lock granularity
both need attention. Per the FreeBSD 8 TCP projects page, I've begun work on
TCP locking and affinity, but I won't have too much substantive progress to
show for a couple of months. I can provide netisr parallel dispatch patches,
a bit dated but probably not hard to update, on request.
So, suffice to say, we are aware that we have more work to do in the area of
parallel work distribution, and by implication, lock granularity in some of
our network stack subsystems. However, I think we're quite well placed to do
it, as we already ship the OS with parallel input queue processing due to
direct dispatch, and have a solid and generally fairly granular locking
implementation.
Robert N M Watson
Computer Laboratory
University of Cambridge
Thanks,
Tiffany.
On 12/29/05, Andre Oppermann <[EMAIL PROTECTED] > wrote:
Markus Oestreicher wrote:
Currently running a few routers on 5-STABLE I have read the
recent changes in the network stack with interest.
You should run 6.0R. It contains many improvements over 5-STABLE.
A few questions come to my mind:
- Can a machine that mainly routes packets between two em(4)
interfaces benefit from a second CPU and SMP kernel? Can both
CPUs process packets from the same interface in parallel?
My testing has shown that a machine can benefit from it but not
much in the forwarding performance. The main benefit is the
prevention of lifelock if you have very high packet loads. The
second CPU on SMP keeps on doing all userland tasks and running
routing protocols. Otherwise your BGP sessions or OSPF hellos
would stop and remove you from the routing cloud.
- From reading the lists it appears that net.isr.direct
and net.ip.fastforwarding are doing similar things. Should
they be used together or rather not?
net.inet.ip.fastforwarding has precedence over net.isr.direct and
enabling both at the same doesn't gain you anything. Fastforwarding
is about 30% faster than all other methods available, including
polling. On my test machine with two em(4) and an AMD Opteron 852
(2.6GHz) I can route 580'000 pps with zero packet loss on -CURRENT.
An upcoming optimization that will go into -CURRENT in the next
few days pushes that to 714'000 pps. Futher optimizations are
underway to make a stock kernel do close to or above 1'000'000 pps
on the same hardware.
--
Andre
_______________________________________________
freebsd-net@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-net
To unsubscribe, send any mail to " [EMAIL PROTECTED]"
_______________________________________________
freebsd-net@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-net
To unsubscribe, send any mail to "[EMAIL PROTECTED]"
_______________________________________________
freebsd-net@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-net
To unsubscribe, send any mail to "[EMAIL PROTECTED]"
