Container to container networking performance degradation

[brunograz]

Moving this issue out of https://github.com/cloudfoundry/cf-networking-release/issues/213 as we have indications that it is related to the Stemcell.

Issue

We currently observe timeouts in C2C when moving CF from Bionic to Jammy. Please note that this issue can only be observed when the Diego cells are migrated from Bionic to Jammy and cannot be reproduced on Bionic stemcells. As additional information, we've also tested in different environments with and without dynamic ASGs.

Steps to Reproduce - See additional information below

• Install cf-deployment [v27.2.0] on Jammy stemcell
• Push two apps and add a network-policy enabling traffic from app A to app B
• cf add-network-policy app-a app-b --protocol tcp --port 8080
• ssh into app-a and try to reach app-b

Expected result

Successful connections from app-a to app-b.

Current result

Sporadic timeouts and slow connections from app-a to app-b.

[backend-wgnnmafs]: Hello!
real    0m2.035s
user    0m0.000s
sys     0m0.007s
[backend-wgnnmafs]: Hello!
real    0m0.018s
user    0m0.000s
sys     0m0.007s
[backend-wgnnmafs]: Hello!
real    0m1.039s
user    0m0.000s
sys     0m0.007s

Workaround

In every CloudFoundry diego cell you should disable a configuration parameter in the networking interface: ethtool -K eth0 tx-udp_tnl-segmentation off && ethtool -K eth0 tx-udp_tnl-csum-segmentation off

This is currently disabled (off) by default on Bionic compared to Jammy.

Further information

Infrastructure: ESXI prepared with NSX-T / NSX-V (tested on both) - not sure if it can be reproduced in other cloud environments.

[cunnie]

I'm having a hard time replicating

My throughput is fine; curl between 2 apps on different diego hosts averages ~0.017 seconds; however, these ESXi hosts do not have the NSX vibs installed and the traffic is passed along a regular VDS (Virtual Distributed Switch) instead of an NSX Segment.

Just to check: you're using the CF default Silk backend, right? Not the NSX-T Container Plug-in, right? If you don't know the answer, you're using the Silk backend.

Technical Details

for app in dora-{0,1}; do
  echo -n "$app: "
  cf ssh $app -c 'echo $CF_INSTANCE_ADDR / $CF_INSTANCE_INTERNAL_IP'
done

Gives me:

dora-0: 10.9.250.18:61000 / 10.255.171.137
dora-1: 10.9.250.17:61000 / 10.255.249.140

cf add-network-policy dora-0 dora-1 --protocol tcp --port 8080
cf ssh dora-0
curl http://10.255.249.140:8080 # Hi, I'm Dora!
time for i in $(seq 1 1000); do curl http://10.255.249.140:8080 > /dev/null 2>&1 ; done

real	0m17.550s
user	0m4.991s
sys	0m6.824s

[brunograz]

Correct, we are using the Silk backend not the NSX-T plug-in. I have tried with different Jammy stemcell versions but only the workaround mentioned above worked as these values are disabled by default in Bionic.

[metron2]

We noticed this as well but didn't get to the solution, just kept an isolation segment of bionic cells for the one team using container-to-container networking.

They deploy an nginx server and use the internal app route to talk to their application (essentially to add authentication).