First we need to build and run the eBPF program:
go generate
go build
sudo ./sockmap
Then we spawn a server and a client process in different network namespaces as well as setting up the cgroup(v2) and the veth pairs using:
mkdir -p /sys/fs/cgroup/unified/test.slice
ip netns add A
ip netns add B
ip -n A link add name veth0 type veth peer name veth0 netns B
ip -n A link set dev lo up
ip -n B link set dev lo up
ip -n A link set dev veth0 up
ip -n B link set dev veth0 up
ip -n A addr add 10.0.0.1/24 dev veth0
ip -n B addr add 10.0.0.2/24 dev veth0
Network namespace A is where we run our server using:
echo $$ > /sys/fs/cgroup/unified/test.slice/cgroup.procs # Put current shell process into the cgroup to which our BPF program is attached to
ip netns exec A sockperf server -i 10.0.0.1 --tcp --daemonize # Run the server inside the network namespace A
And then run the client (new shell) in the network namespace B using:
echo $$ > /sys/fs/cgroup/unified/test.slice/cgroup.procs # Put current shell process into the cgroup to which our BPF program is attached to
ip netns exec B sockperf ping-pong -i 10.0.0.1 --tcp --time 30 # Run the client inside the network namespace B
The latency results I get on my machine is 18.282 usec without the Sockmap, and with the latency reduces to 12.997 usec, which is a aprox. 30% improvement!