On Mon, Sep 01, 2014 at 08:08:25PM +0200, Tom Ivar Helbekkmo wrote: > Michael van Elst <[email protected]> writes: > > > This doesn't necessarily mean that the buffer space is too small. It > > often also means that you try to send data faster than possible > > through the particular hardware. > > [...] > > It could be that the kernel address space is exhausted but this > > shouldn't really happen with default settings. > > This rung a bell, and got me started looking for things that might be > other than default. I found some stuff in /etc/sysctl.conf picked up > from https://wiki.netbsd.org/tutorials/tuning_netbsd_for_performance/ > a long time ago, when the machine had quite different tasks: > > net.inet.tcp.recvbuf_auto=1 > net.inet.tcp.sendbuf_auto=1 > net.inet.tcp.sendbuf_max=16777216 > net.inet.tcp.recvbuf_max=16777216 > vm.filemax=20 > > Reverting those to their defaults solved my problem.
Note that scaling up the send/recv buf alone won't help without also raising kern.somax. Joerg
