[Intel-wired-lan] X710 disable flow hashing?

Ray Bellis ray at isc.org
Thu Feb 1 09:10:35 UTC 2018


On 31/01/2018 17:46, Alexander Duyck wrote:

> Generally what I have found works for me is to use Flow Director rules
> based on the source/destination IP and source/destination port numbers
> and then have my traffic generator alter either the source IP or
> source port for the traffic in order to get the round robin effect
> without necessarily having to modify the driver. For example pktgen
> supports something like this by specifying a port range when
> generating a flow.

The traffic generator I'm using can only produce a very limited range of
source port numbers (24), which is insufficient to produce an even
distribution in the queues with flow hashing.

In fact, with that small a range I consistently get _no traffic at all_
to one of my 12 queues.   See for example this one second snapshot of
the per-queue RX rates:

   0      25856
   1      50669
   2      37624
   3          0
   4      35778
   5      13002
   6      24228
   7      35615
   8      24812
   9      22804
  10      26047
  11      50831

I suppose that with only 24 source ports I could perhaps create 24
individual flow director rules, though?

> This is a pretty old version of the driver. Does it need to be that
> specific driver or could you work with a later version such as the one
> provided by e1000.sf.net.  If driver only modifications don't concern
> you then probably the e1000.sf.net driver might be the best way to go
> since all you need to do is get it to a functional version that meets
> your needs and then you could probably just lock in on that version
> for all future tests.

For now, it needs to be that version.  See threads from November '16
where I reported a significant 50% regression in PPS rates when
upgrading beyond 1.3.9.

> I assume you must be benchmarking something up at the user-space
> layer. Hopefully my suggestions work for you, if not let me know.

Yes, I'm benchmarking DNS server software.

(https://www.isc.org/blogs/isc-performance-lab/)

kind regards,

Ray



More information about the Intel-wired-lan mailing list