On Mon, Apr 24, 2017 at 06:26:43PM -0400, David Miller wrote: > From: Jesper Dangaard Brouer <brouer@xxxxxxxxxx> > Date: Mon, 24 Apr 2017 16:24:05 +0200 > > > I've done a very detailed evaluation of this patch, and I've created a > > blogpost like report here: > > > > https://prototype-kernel.readthedocs.io/en/latest/blogposts/xdp25_eval_generic_xdp_tx.html > > Thanks for doing this Jesper. Yes, this is excellent. I'm not all the way thru it, but I looked at the data and corroborate the results you are seeing. My results for both optimized and generic XDP for xdp_bench01_mem_access_cost --action XDP_DROP --readmem are quite similar to yours (11.7Mpps and 7.8Mpps, respectively for me 11.7Mpps and 8.4Mpps for you). I also noted (as you did) that there is no discernible difference running xdp_bench01_mem_access_cost with or without the --readmem option since the packet data is already being accessed that late it the stack. > > > I didn't evaluate the adjust_head part, so I hope Andy is still > > planning to validate that part? > > I was hoping he would post some results today as well. > > Andy, how goes it? :) Sorry for the delayed response. I was AFK yesterday, but based on testing from Friday and what I wrapped up today all looks good to me. On my system (i7-6700 CPU @ 3.40GHz) the reported and actual TX throughput for xdp_tx_iptunnel is 4.6Mpps for the optimized XDP. For generic XDP the reported throughput of xdp_tx_iptunnel is 4.6Mpps but only ~880kpps actually on the wire. It seems to me that can be fixed with a follow-up for offending drivers or the stack if deemed that there is a real error there. > Once the basic patch is ready and integrated in we can try to do > xmit_more in generic XDP and see what that does for XDP_TX > performance. Agreed.
|