---------- Forwarded message --------- From: Kálmán Kiss <[email protected]> Date: Sun, Nov 2, 2025 at 10:17 PM Subject: Re: Slow 'real world' network performance To: Peter Miller <[email protected]>
I got curious that tuning those sysctl params dont yield any real improvements even in my very simple network setup (no congestion seen). Playing on an amd64 windows host <-> netbsd in a virtualbox env, I get a 30-50% increase in average download speed, transfering a big file through ssh/sftp from the vm by using a custom kernel with doubled mbuf size (1024) and mclshift (12) value (param.h). My only other modifications are maxusers 128 in the kernel config, and setting cubic congestion algo in sysctl. I cant say I fully understand what Im doing, but for me it seems the mbuf params should be configurable at boot or with sysctl? :) Anyway, I plan to test this customizations in my home network with various clinent hosts and OSes. On Wed, Oct 29, 2025 at 2:45 AM Peter Miller <[email protected]> wrote: > On Sat, Oct 25, 2025 at 4:09 PM Michael van Elst <[email protected]> > wrote: > > The examples given here are already all you can tune. Basically you > > bump the buffer sizes to allow for larger TCP windows. > ... > > This is the major reason for your observations (and shows that > > NetBSD has room for improvement). > > Thank you, good to know. > > Since I get good speeds with FreeBSD I figure the sysctl settings > ought to be a good reference. Hopefully I got the right sysctls since > the naming is a bit different. > > > ... > > Here's the /etc/sysctl.conf I tested. > > kern.sbmax=2097152 > kern.somaxkva=536870912 > net.inet.tcp.sendbuf_max=2097152 > net.inet.tcp.recvbuf_max=2097152 > > net.inet.tcp.init_win=10 > net.inet.tcp.init_win_local=10 > net.inet.tcp.congctl.selected=cubic > net.inet.tcp.delack_ticks=40 > > hw.vioif0.tx.process_limit=512 > hw.vioif0.rx.process_limit=512 > > And I got very similar results. The biggest difference by far is > network congestion etc. that's out of my control. > > Thanks for the help. > > -- > Thanks > Peter >
