[ovs-dev] [PATCH 01/11] dpif: Pool dpif_recv() calls.

[Rajahalme, Jarno (NSN - FI/Espoo)]

On Feb 12, 2013, at 1:54 , ext Ben Pfaff wrote:

> On Mon, Feb 11, 2013 at 04:46:17PM +0200, Jarno Rajahalme wrote:
>> Take ofproto-dpif upcall recv pooling down to the system call interface.
>> 
>> Signed-off-by: Jarno Rajahalme <jarno.rajahalme at nsn.com>
> 
> I tried this out with my test case.  I found that it yields a small
> performance loss of about 3% with flow tables that just contain a
> "normal" action and about the same with complicated flow tables that
> contain multiple levels of resubmit.
> 
> I'm surprised that it produces such dramatically better results for
> your test case.

So was I. I expected some benefit only after patch 02/11 (recvmmsg()).

> 
> Our test cases are quite different, though.  Yours has a constant rate
> at the source, and measures the fraction of packets that makes it
> through.  Mine essentially never drops a packet because it only sends
> a new packet when the reply to a previous one has been received.  I am
> not certain which is a better model for actual network behavior.  Do
> you have any thoughts


I designed the test case to show the difference in throughput on upcall processing path under heavy load. In this case any CPU savings would show up as increased throughput. Your test case is probably more sensitive to latency, which I did not specifically look for.

Based on your result I would think there might be only a few packets coming up at a time, where the small set-up cost for batching would not be amortized. In this case small latency increase could show in your test.  One way to test for this is to temporarily reduce the batch size (in ofproto_dpif), and see if that makes any difference.

It may also be that there is some unnecessary overhead causing extra latency, I'll see if I can find something.

  Jarno