On Sun, Jan 24, 2016 at 01:22:20PM -0800, Jonathon Sisson wrote: > On Sun, Jan 24, 2016 at 09:08:32PM +0100, Mike Belopuhov wrote: > > I haven't seen that on my test box (not AWS), but maybe reverting > > the minimum number of rx slots back to 32 can help? > > > > http://cvsweb.openbsd.org/cgi-bin/cvsweb/src/sys/dev/pv/if_xnf.c.diff?r1=1.9&r2=1.10 > > > Reverting to 32 fixed the dhcp issue. > > I'll go ahead and get those dmesgs for you now =) > > Thanks again! > Mike,
A series of quick iperf tests showed the following: iperf server: iperf -s (Amazon Linux) iperf client: iperf -c $SRVIP -dt 300 -i 30 Amazon Linux <-> Amazon Linux (same AZ/VPC subnet) ~690 Mbits. (M3.larges) OpenBSD-CURRENT <-> Amazon Linux (same AZ/VPC subnet) ~400 Mbits. (M3.larges again) Each test, I ran the same Amazon Linux instance as the server. Since I was running bi-directional, I did notice that the OpenBSD machine accepting incoming traffic was slower than it sending traffic: [ 4] 0.0-30.0 sec 1.48 GBytes 422 Mbits/sec [ 5] 0.0-30.0 sec 1.08 GBytes 310 Mbits/sec I chose M3 due to the lack of support for SRIOV, so the Amazon Linux instances wouldn't utilize it: [root@ip-172-31-46-242 ~]# ethtool -i eth0 | grep driver driver: vif I'm gathering the dmesgs and will post links when I have them uploaded. -Jonathon
