Re: Performance impact of PF on APU2

Not sure what do do know. Should I open a bug for more visibility?

To be honest, my WAN connection is way lower than the max
measured here with CURRENT, but I don´t want to discover when
upgrading to 6.5 that I lost 40% percent of performance again.

I would be more than happy to help with the investigations
(what/where to look, what setting to play with).

Otherwise I will have to switch to another OS, and I would rather
not. (Simple NAT rules with FreeBSD 11.2: ~890Mbits/s, with
OpenWRT ~950Mbits/s)

I am also surprised to see that using the APU2 as an iperf3 client
cannot max a gigabit connection (without pf involved).

I get that performance is not the main focus on OpenBSD, but
this regression is kind of scary to me.

Thanks,

On Thu, 04 Oct 2018 17:33:37 -0700
Benjamin Petit <bpetit@yahoo.fr> wrote:

> I am very brave indeed :)
>
> OpenBSD 6.4 (GENERIC.MP) #0: Wed Oct 3 13:49:29 CEST 2018
> hrvoje@r620-2.srce.hr:/sys/arch/amd64/compile/GENERIC.MP
> real mem = 1996279808 (1903MB)
> avail mem = 1926565888 (1837MB)
> mpath0 at root
> scsibus0 at mpath0: 256 targets
> mainbus0 at root
> bios0 at mainbus0: SMBIOS rev. 2.7 @ 0x77fd7020 (7 entries)
> bios0: vendor coreboot version "v4.0.19" date 20180902
> bios0: PC Engines apu2
>
> But I see even worst performance now: 458 Mbits/sec
>
>
> On Thu, 2018-10-04 at 22:26 +0200, Hrvoje Popovski wrote:
> > On 4.10.2018. 5:58, Benjamin Petit wrote:
> > > Ok so I compared 6.3-release, 6.3-release+syspatches(=stable?) and
> > > the latest snapshot from October 2.
> > >
> > > I measured iperf3 throughput between A and B, like this:
> > > PC A <---> APU2 <---> PC B
> > >
> > > pf rules are the one shipped by default in 6.3:
> > >
> > > gw# pfctl
> > > -sr
> > >
> > > block return all
> > > pass all flags S/SA
> > > block return in on ! lo0 proto tcp from any to any port 6000:6010
> > > block return out log proto tcp all user = 55
> > > block return out log proto udp all user = 55
> > >
> > > OpenBSD 6.3 RELEASE:
> > > - pf enabled: 841 Mbits/sec
> > > - pf disabled: 935 Mbits/sec
> > >
> > > OpenBSD 6.3 + Syspatch:
> > > - pf enabled: 803 Mbits/sec
> > > - pf disabled: 936 Mbits/sec
> > >
> > > OpenBSD CURRENT:
> > > - pf enabled: 526 Mbits/sec (541 with kern.pool_debug=0)
> > > - pf disabled: 934 Mbits/sec
> > >
> > > So there is a small perf drop when applying all syspatches to 6.3
> > > (not sure which one cause the drop),
> > > but the performance drop SIGNIFICANTLY using the latest snapshot.
> > >
> > > Am I missing something? (I really hope I am)
> > >
> >
> > Hi,
> >
> > if you're feeling brave enough and you can test/experiment
> > with pf you can download openbsd kernel with experimental MP support
> > from here http://kosjenka.srce.hr/~hrvoje/zaprocvat/smpfbsd
> >
> > SHA256 (smpfbsd) =
> > e95e94190a0e52de7690b3278cfab14985817089e7a53615cd2599420593b32c
> >
> > this kernel is compiled with option WITH_PF_LOCK and NET_TASKQ=4
> >
> > before you download it please backup your active kernel so if
> > something
> > goes wrong you can put it back ..
> >
> > cp /bsd /goodbsd
> > cp smpfbsd /bsd
> > reboot
> >
> > if something goes wrong at boot prompt before kernel starts to boot
> > you
> > can boot old kernel with command - boot goodbsd
> >
> > i'm running this kernel for few days and i'm hitting pf, pfsync and
> > pflow quite hard and it seems stable :)
> >
>


--
Benjamin Petit <bpetit@yahoo.fr>