[Oisf-users] Packet drops on core 0 when file extraction is enabled?

[Eric Leblond]

Hello,

On Thu, 2014-01-30 at 12:14 -0800, Cooper F. Nelson wrote:
> Hi all,
> 
> I'm experimenting with file extraction again.
> 
> It's working well with the exception that core 0 is dropping lots of
> packets (using worker mode).  I have management threads set to use all
> cores, low priority. 

Doing this was needed. Maybe you can improve things a little bit by not
using core 0 for management threads.

>  Is there any setting available to mitigate this?

Intel card with ixgbe driver are using RSS queue 0 to send most crap
packets. For example, UDP are sent to RSS 0 by default if you did not do
	sudo ethtool -n eth3 rx-flow-hash udp4
It is also sending to queue 0 all packets that can not be load balanced
via the embedded hash function. For example, IPv6 fragmented packet fall
into that category.

So core 0 receives the bad traffic and that could explain why it has
more drop.


> It seems as if some of the file tracking is single-threaded, or
> otherwise tied to core 0.  Example below; with file tracking disabled
> packet loss is the same for all cores.
> 
> > capture.kernel_packets    | AFPacketeth21             | 34950994
> > capture.kernel_drops      | AFPacketeth21             | 11398893
> > capture.kernel_packets    | AFPacketeth22             | 39400674
> > capture.kernel_drops      | AFPacketeth22             | 63314

It could be interesting to compare the work done on CPU 0 and on the
other CPUs. You can use linux-tool for that by running perf top. To get
info on core 0:
	perf top -C 0
Then compare with another CPU:
	perf top -C 1

Regarding your point about file tracking. Are you doing file
extraction ? If that's the case maybe there is interference with output
device (maybe check interrupts to see if CPU 0 is handling the block
device).

BR,
-- 
Eric Leblond <eric at regit.org>