On 6/28/12 3:12 PM, Sašo Kiselkov wrote:
On 06/28/2012 10:36 AM, Garrett D'Amore wrote:
I haven't looked at the code, but it wouldn't surprise me if multicast fanout sucked.
Multicast traffic is generally considered by *many* nic drivers and IP stacks to be lower
bandwidth than unicast (a *lot* lower), and often gets "slowpath" treatment.
Its a little surprising to me that you have enough multicast traffic to justify running
in polled mode.
Upon further investigation, I really seem to be hitting bug #918.
Trouble is, even without VLANs configured, I'm still hitting this thing,
so I'm kind of puzzled as to why, however, I can clearly see
mac_rx_srs_long_fanout being entered into from mac_rx_srs_fanout with
IPv4 packets and, predictably, the fanout routine selects the first CPU
in the CPU list to handle them, so that ends up pouring all of the load
onto a single CPU.
I see that this bug has been open for over a year and hasn't been
touched in 5 months... Can I help to somehow fix it? Any pointers on how
to compile a custom mac module and load it into the kernel? (I've not
yet compiled kernel stuff for Illumos, only for Linux, but I'd be very
happy to learn it.) Where should I start reading? Thanks!
Joyent fixed this when the bug was filed. Dan McDonald was looking at
upstreaming it, as he's the assignee. I don't know what happened to it
from there.
To build a custom module, you'll want to start by building illumos
itself. You'll probably want to get started off of the wiki page:
http://wiki.illumos.org/display/illumos/How+To+Build+illumos.
The best way to update and test the changes depend on what distribution
of illumos you're using.
Robert
-------------------------------------------
illumos-discuss
Archives: https://www.listbox.com/member/archive/182180/=now
RSS Feed: https://www.listbox.com/member/archive/rss/182180/21175430-2e6923be
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21175430&id_secret=21175430-6a77cda4
Powered by Listbox: http://www.listbox.com
