Skip site navigation (1)Skip section navigation (2)
Date:      Sat, 14 Sep 2024 11:25:16 +0100
From:      Sad Clouds <cryintothebluesky@gmail.com>
To:        Zhenlei Huang <zlei@FreeBSD.org>
Cc:        Mark Saad <nonesuch@longcount.org>, FreeBSD Net <freebsd-net@freebsd.org>
Subject:   Re: Performance issues with vnet jails + epair + bridge
Message-ID:  <20240914112516.cfb31bae68ab90b83ca7ad4b@gmail.com>
In-Reply-To: <A95066A8-F5FC-451B-85CE-C463952ABADE@FreeBSD.org>
References:  <20240913100938.3eac55c9fbd976fa72d58bb5@gmail.com> <39B2C95D-1E4F-4133-8923-AD305DFA9435@longcount.org> <20240913155439.1e171a88bd01ce9b97558a90@gmail.com> <A95066A8-F5FC-451B-85CE-C463952ABADE@FreeBSD.org>

next in thread | previous in thread | raw e-mail | index | archive | help
On Sat, 14 Sep 2024 10:45:03 +0800
Zhenlei Huang <zlei@FreeBSD.org> wrote:

> The overhead of vnet jail should be neglectable, compared to legacy jail
> or no-jail. Bare in mind when VIMAGE option is enabled, there is a default
> vnet 0. It is not visible via jls and can not be destroyed. So when you see
> bottlenecks, for example this case, it is mostly caused by other components
> such as if_epair, but not the vnet jail itself.

Perhaps this needs a correction - the vnet itself may be OK, but due to
a single physical NIC on this appliance, I cannot use vnet jails
without virtualised devices like if_epair(4) and if_bridge(4). I think
there may be other scalability bottlenecks.

I have a similar setup on Solaris

Here devel is a Solaris zone with exclusive IP configuration, which I
think may be similar to FreeBSD vnet. It has a virtual NIC devel/net0
which operates over the physical NIC also called net0 in the global
zone:

$ dladm
LINK                CLASS     MTU    STATE    OVER
net0                phys      1500   up       --
net1                phys      1500   up       --
net2                phys      1500   up       --
net3                phys      1500   up       --
pkgsrc/net0         vnic      1500   up       net0
devel/net0          vnic      1500   up       net0

If I run TCP bulk data benchmark with 64 concurrent threads, 32
threads with server process in the global zone and 32 threads with
client process in the devel zone, then the system evenly spreads the
load across all CPU cores and none of them are sitting idle:

$ mpstat -A core 1
 COR minf mjf xcal  intr ithr  csw icsw migr smtx  srw  syscl  usr sys  st idl sze
   0    0   0 2262  2561    4 4744 2085  209 7271    0 747842  272 528   0   0   8
   1    0   0 3187  4209    2 9102 3768  514 10605   0 597012  221 579   0   0   8
   2    0   0 2091  3251    7 6768 2884  307 9557    0 658124  244 556   0   0   8
   3    0   0 1745  1786   16 3494 1520  176 8847    0 746373  273 527   0   0   8
   4    0   0 2797  2767    3 5908 2414  371 7849    0 692873  253 547   0   0   8
   5    0   0 2782  2359    5 4857 2012  324 9431    0 684840  251 549   0   0   8
   6    0   0 4324  4133    0 9138 3592  538 12525   0 516342  191 609   0   0   8
   7    0   0 2180  3249    0 6960 2926  321 8825    0 697861  257 543   0   0   8

With FreeBSD I tried "options RSS" and increasing "net.isr.maxthreads"
however this resulted in some really flaky kernel behavior. So I'm
thinking that if_epair(4) may be OK for some low-bandwidth use cases,
i.e. testing firewall rules, etc, but not suitable for things like
file/object storage servers, etc.



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20240914112516.cfb31bae68ab90b83ca7ad4b>