Hi Joo,
On 24.08.2017 03:30, Joo Kim wrote:
I mean, is the path of vr_interface_common_hook() receiving packets via
SoftIrq(NAPI poll) OK from performance perspective? I remember
As far as I understand, NAPI is the best way (performance-wise) to get
packets from the NIC. This being said, it's still interrupt-driven and
(relatively) slow, that's why DPDK was born - but that's the whole
another story.
Registering the hook to inject packets into vSwitch/vRouter seems a
de-facto standard way as well. Open vSwitch does the same in its
in-kernel datapath, for instance.
ksoftirqd is easily reaching high 90 percent when pumping up high rate
traffic into Linux stack
Perhaps you want to look at Receive Packet Steering (RPS) feature of the
Linux kernel to spread the netif_receieve_skb() load across your cores.
HTH,
Valentine
On Tue, Aug 15, 2017 at 9:46 PM, Joo Kim <itsolut...@gmail.com
<mailto:itsolut...@gmail.com>> wrote:
Hello,
Our host OS is Centos7-based.
And l happened to see this thread:
http://lists.opencontrail.org/pipermail/dev_lists.opencontrail.org/2015-January/001844.html
<http://lists.opencontrail.org/pipermail/dev_lists.opencontrail.org/2015-January/001844.html>
My query:
Does Vrouter change anything in the packet path code/module of
receiving packets via SoftIrq(NAPI poll) + netif_receive_skb()?
Looks like not.
So, prior to measuring VRouter performance on my setup, if the
packet path of receiving packets via SoftIrq(NAPI poll) +
netif_receive_skb() appears to be not that fast, then do I need to
tune my system based on the tips suggested in above mailing list
thread?
Thanks
_______________________________________________
Dev mailing list
Dev@lists.opencontrail.org
http://lists.opencontrail.org/mailman/listinfo/dev_lists.opencontrail.org
_______________________________________________
Dev mailing list
Dev@lists.opencontrail.org
http://lists.opencontrail.org/mailman/listinfo/dev_lists.opencontrail.org
