Forward memory explode after 24h of running

[denis-tingaikin]

Steps to reproduce

1. Run the forwarder 24-48h

Actual: memory explode happens and it restarts
Expected: forwarder should be stable.

Details

Reproducing with releases: v1.12.0, v1.13.0, v1.13.1

forwarder-vpp_goroutineprofiles_20240706091528.tar.gz
forwarder-vpp_memprofiles_20240706091613.tar.gz

[ljkiraly]

Now the problem occured after 8 hours of testing.
~87 MB memory increase during 1 hour:

go routine profiles:
forwarder-vpp_goroutineprofiles_20240711170319.tar.gz
memory profiles:
forwarder-vpp_memprofiles_20240711170404.tar.gz

Preparing the logs.

[denis-tingaikin]

Yeah, it looks the same as after 24+h of running. Waiting logs.

[ljkiraly]

Forwarder logs for the last occurrence
forwarder-vpp-mem-explosion-logs-20240711-1.tar.gz

[denis-tingaikin]

Could we also get the mapping of pods to nodes?

[denis-tingaikin]

as I can see from logs forwarder-vpp-qqpfp is located on the node n4, right?

[denis-tingaikin]

@ljkiraly I've prepared a fix that should improve the situation in times tinden/cmd-forwarder-vpp:v1.13.2-fix.1. If it works then we'll need to run a few more tests to get a more effective solution.

[ljkiraly]

as I can see from logs forwarder-vpp-qqpfp is located on the node n4, right?

Yes.

forwarder-vpp-56726 1/1 Running 0 9h 10.0.10.106 pool1-ceph-seliics06744-n7
forwarder-vpp-72kmm 1/1 Running 0 9h 10.0.10.103 pool1-ceph-seliics06746-n5
forwarder-vpp-9k6w4 1/1 Running 0 9h 10.0.10.101 control-plane-control-seliics06748-n1
forwarder-vpp-llbts 1/1 Running 0 9h 10.0.10.102 control-plane-control-seliics06750-n3
forwarder-vpp-qqpfp 1/1 Running 0 9h 10.0.10.104 pool1-ceph-seliics06745-n4
forwarder-vpp-tnsd9 1/1 Running 0 9h 10.0.10.105 pool1-ceph-seliics06747-n6
forwarder-vpp-wsx8x 1/1 Running 0 9h 10.0.10.100 control-plane-control-seliics06749-n2

[ljkiraly]

@denis-tingaikin The test now failed after 16 hours with tinden/cmd-forwarder-vpp:v1.13.2-fix.1
This time no server listening on forwarder pod so no profiles can be collected, but find the logs attached.

logs-2024-07-13-09-15-33.tar.gz

[denis-tingaikin]

@ljkiraly OK, we have some small differences here.

For the previous patch in 1h, we get ~87 MB.
For the last patch in 1 hour, we get ~50 MB.

(It might be just luck, but let's keep an eye on it.)

I also prepared a new version of forwarder-vpp: tinden/cmd-forwarder-vpp:v1.13.2-fix.3 

The profile server is available.

Question: Is it possible to get forwarded logs from the forwarders?

Means run something like this for all forwarders (before test start):

kubectl logs -f -n $namespace $forwarderName1 > fwd1.txt
kubectl logs -f -n $namespace $forwarderName2 > fwd2.txt

It could show more details.