On 2014/11/06 19:55, Jonathan Gray wrote: > Here is what I see connecting from a bbb on nfs root via > a 100Mb switch to another host: > > tcpbench > tcp 5.2 Mbps > udp 11.6 Mbps 990 pps > > iperf > tcp 5.29 Mbits/sec > udp 1.05 Mbits/sec > > I'm not sure why udp differs here.
By default, iperf limits bandwidth to 1Mb/s for UDP, this is probably a good idea as a default as it prevents problems with uncareful use. Accidentally running tcpbench -u on a high bandwidth connection can be exciting ;-) You can use e.g. "-b 100M" to let it run faster. > iperf > tcp 92.5 Mbits/sec > udp 1.05 Mbits/sec > > Perhaps the switch frontend/ALE configuration is not set > up correctly? I'm sure you must have read a bit about that > when looking into adding support for the other port. > > It looks like there is not a clean seperation of registers/interrupts/dma > for the different ports which is annoying. > > The driver in OpenBSD does not attempt to do interrupt mitigation either. > > Though perhaps with such a large difference something more fundamental > like the processor cache/pmap/frequency handling could be implicated? Network i/o has always seemed fairly slow on OpenBSD arm platforms compared to other OS running on the hardware.
