Re: [ovs-dev] [RFC PATCH 4/7] net: openvswitch: ovs_vport_receive reduce stack usage

["Nicholas Piggin" <npiggin@gmail.com>]

On Fri Sep 29, 2023 at 1:26 AM AEST, Aaron Conole wrote:
> Nicholas Piggin <npiggin@gmail.com> writes:
>
> > Dynamically allocating the sw_flow_key reduces stack usage of
> > ovs_vport_receive from 544 bytes to 64 bytes at the cost of
> > another GFP_ATOMIC allocation in the receive path.
> >
> > XXX: is this a problem with memory reserves if ovs is in a
> > memory reclaim path, or since we have a skb allocated, is it
> > okay to use some GFP_ATOMIC reserves?
> >
> > Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
> > ---
>
> This represents a fairly large performance hit.  Just my own quick
> testing on a system using two netns, iperf3, and simple forwarding rules
> shows between 2.5% and 4% performance reduction on x86-64.  Note that it
> is a simple case, and doesn't involve a more involved scenario like
> multiple bridges, tunnels, and internal ports.  I suspect such cases
> will see even bigger hit.
>
> I don't know the impact of the other changes, but just an FYI that the
> performance impact of this change is extremely noticeable on x86
> platform.
>
> ----
> ip netns add left
> ip netns add right
>
> ip link add eth0 type veth peer name l0
> ip link set eth0 netns left
> ip netns exec left ip addr add 172.31.110.1/24 dev eth0
> ip netns exec left ip link set eth0 up
> ip link set l0 up
>
> ip link add eth0 type veth peer name r0
> ip link set eth0 netns right
> ip netns exec right ip addr add 172.31.110.2/24 dev eth0
> ip netns exec right ip link set eth0 up
> ip link set r0 up
>
> python3 ovs-dpctl.py add-dp br0
> python3 ovs-dpctl.py add-if br0 l0
> python3 ovs-dpctl.py add-if br0 r0
>
> python3 ovs-dpctl.py add-flow \
>   br0 'in_port(1),eth(),eth_type(0x806),arp()' 2
>   
> python3 ovs-dpctl.py add-flow \
>   br0 'in_port(2),eth(),eth_type(0x806),arp()' 1
>
> python3 ovs-dpctl.py add-flow \
>   br0 'in_port(1),eth(),eth_type(0x800),ipv4()' 2
>
> python3 ovs-dpctl.py add-flow \
>   br0 'in_port(2),eth(),eth_type(0x800),ipv4()' 1
>
> ----
>
> ex results without this patch:
> [root@wsfd-netdev60 ~]# ip netns exec right ./git/iperf/src/iperf3 -c 172.31.110.1
> ...
> [  5]   0.00-10.00  sec  46.7 GBytes  40.2 Gbits/sec    0             sender
> [  5]   0.00-10.00  sec  46.7 GBytes  40.2 Gbits/sec                  receiver
>
>
> ex results with this patch:
> [root@wsfd-netdev60 ~]# ip netns exec right ./git/iperf/src/iperf3 -c 172.31.110.1
> ...
> [  5]   0.00-10.00  sec  44.9 GBytes  38.6 Gbits/sec    0             sender
> [  5]   0.00-10.00  sec  44.9 GBytes  38.6 Gbits/sec                  receiver
>
> I did testing with udp at various bandwidths and this tcp testing.

Thanks for the test case. It works perfectly in the end, but it took me
days to get there because of a random conspiracy of problems I hit :(
Sorry for the slow reply, but I was now able to test another idea for
this. Performance seems to be within the noise with the full series, but
my system is only getting ~half the rate of yours so you might see more
movement.

Instead of slab it reuses the per-cpu actions key allocator here.

https://github.com/torvalds/linux/commit/878f01f04ca858e445ff4b4c64351a25bb8399e3

Pushed the series to kvm branch of https://github.com/npiggin/linux

I can repost the series as a second RFC but will wait for thoughts on
this approach.

Thanks,
Nick