Comment 15 for bug 1281366
Revision history for this message
| Rafael David Tinoco (rafaeldtinoco) wrote | #15 |
© 2004
Canonical Ltd.
•
Terms of use
•
Data privacy
•
Contact Launchpad Support
•
Blog
•
Careers
•
System status
•
1b1ed1a
(Get the code!)
Bryan
Id like you to reconsider this test if possible. Better results have also been confirmed by others (including customers).
Fake routers creation performance depends on other issues (not only covered by this bug) like CPU scalability (since netlink generated work is delayed by kernel and offloaded to RCU callbacks). There are some other issues being investigated for fake router creation but for this specific bug I can confirm results several times.
Keep in mind that results are sometimes less then 10% better because of kernel RCU offload characteristics (like for example when using only 1 CPU for the test).
My testcase:
I'm using last good commit generated kernel for this case (34ed6246). After this specific commit (upstream kernel) performance for fake router creation brakes (I'll put other BUG information and # at the end of this case, when appropriate).
By now, from kernel 3.2 to kernel 3.9 (up to this commit), general behavior is to have RCU callbacks enabled (no configuration on what cpu to offload up to this point) and this is what "NOCB" stands for on the attached file.
Independently of the kernel (running on precise), iproute userland tool + this patch gives better results.
Please take a look at the attached file specifically on "average" fields.
Thanks in advance.