On Fri, 6 Dec 2024, Michael van Elst wrote:
On Fri, Dec 06, 2024 at 02:49:03PM +0000, Stephen Borrill wrote:At 13:17 I doubled the values again: 13:17 sysctl -w kern.somaxkva=134217728 13:17 sysctl -w kern.sbmax=134217728 The warnings continued:sbmax is the limit for a single socket buffer. somaxkva is the limit for the whole network stack. If you make the same size, this allows one socket buffer to eat all allowed memory. I doubt that this is the problem, otherwise you'd need more mbufs allocated to fill the larger somaxkva now, but the netstat output didn't change. But please verify with somaxkva >> sbmax (and a much smaller sbmax).
I'm not sure I got that figure from and my recvbuf/sendbuf settings. Various posts on the mailing lists, I guess. My local commit message for those settings was "TCP tuning to improve scp and rsync-over-ssh performance", but I didn't bother giving a reference unfortunately. It probably relates to my previous complaints about terrible scp performance with a NetBSD ssh server compared to Linux (both with a NetBSD ssh client).
Here's what I am currently using after reducing sbmax: net.inet.tcp.recvbuf_max=16777216 net.inet.tcp.recvbuf_inc=786432 net.inet.tcp.sendbuf_max=16777216 net.inet.tcp.sendbuf_inc=786432 net.inet.tcp.recvspace=262144 net.inet.tcp.sendspace=262144 kern.sbmax=16777216 kern.somaxkva=268435456 There won't be any load until Monday now. -- Stephen