Andrew Gallatin writes: > xmit routine was called 683441 times. This means that the queue was > only a little over two packets deep on average, and vmstat shows idle > time. I've tried piping additional packets to nghook mx0:orphans > input, but that does not seem to increase the queue depth. >
The problem here seems to be that rather than just slapping the packets onto the driver's queue, ng_source passes the mbuf down to more of netgraph, where there is at least one spinlock, and the driver's ifq lock is taken and released a zillion times by ether_output_frame(), etc. A quick hack (appended) to just slap the mbufs onto the if_snd queue gets me from ~410Kpps to 1020Kpps. I also see very deep queues with this (because I'm slamming 4K pkts onto the queue at once..). This is nearly identical to the linux pktgen figure on the same hardware, which makes me feel comfortable that there is a lot of headroom in the driver/firmware API and I'm not botching something in the FreeBSD driver. BTW, did you see your 800Kpps on 4.x or 5.x? If it was 4.x, what do you see on 5.x if you still have the same setup handy? Thanks, Drew
ng_source.diff
Description: diff to make ng_source skip netgraph detour
_______________________________________________ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-net To unsubscribe, send any mail to "[EMAIL PROTECTED]"
