Add a delay between killing teamd processes #3325
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
When killing 10 or more teamd processes, add a delay of 0.1 seconds after every 10 kill signals/proceses. This is because in the LAG scale tests (in `ecmp/inner_hashing/test_inner_hashing_lag.py` in sonic-mgmt), it may create 100 LAGs, and when destroying them all, some of those LAGs may fail to be properly destroyed, leaving some stale port channels around. This seems to be because the netlink socket buffers on which the teamd processes get notifications become full with events of the other port channels/interfaces going down. As a workaround, add some delays in killing the teamd processes, so that the netlink buffers don't become full, causing messages to get dropped. This delay was randomly chosen, and it seems to work well with 100 LAGs on a KVM. It can probably made to be a bit more aggressive if needed (i.e. maybe 0.05 seconds every 20 processes). Signed-off-by: Saikrishna Arcot <[email protected]>
Signed-off-by: Saikrishna Arcot <[email protected]>
This requires overriding some libc functions and capturing information about kill signals sent or intercepting file open operations. Signe -off-by: Saikrishna Arcot <[email protected]>
|
/azpw run |
|
/AzurePipelines run |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
Comparing the time needed to send SIGTERM to the teamd processes before and after this change, it appears that the time is roughly the same for about 70 LAGs, as tested on a physical device. Before: After: In both cases, for 70 LAGs, it took about 0.6-0.7 seconds to send SIGTERM to the teamd processes, but the distribution of SIGTERMs sent is different. However, on this device, there are still some netlink messages getting dropped resulting in the cleanup not being complete. |
|
@saiarcot895 Can you please run your changes with test_po_cleanup and test_po_cleanup_after_reload? We are noticing these tests statically fail with your changes |
100ms might not be enough on slow systems for the teamd shutdown sequence to actually be staggered. Signed-off-by: Saikrishna Arcot <[email protected]>
… into teamd-delay-kill
|
@saiarcot895 Do we have an updated fix? |
|
Updated syslog after changes: SIGTERM signal to the teamd supervisord was sent at 07:41:40, and the container exited at 07:41:46. Teamd container is also configured to have a 60-second wait time (instead of the 10-second wait time) for termination, meaning docker will wait 60 seconds for the container to go down before forcefully killing it. This means that a 6-second wait time here is fine. |
|
/azp run |
|
Azure Pipelines successfully started running 1 pipeline(s). |
judyjoseph
reviewed
Dec 15, 2024
cfgmgr/teammgr.cpp
Outdated
| pid_t pid; | ||
| if (++sleepCounter % 10 == 0) { | ||
| // Sleep for 100 milliseconds so as to not overwhelm the netlink |
There was a problem hiding this comment.
Do we need to change comment .. is it 100ms or 200ms
There was a problem hiding this comment.
Also instead of giving a one shot 200ms delay .. can we put 10ms for each delete ? The issues we have seen before is when we delete a large number of Portchannels few stale remains.
So now here you are optimizing the call from shell/cmd approach to system_call() approach which will make it execute faster
I felt the older shell/cmd way of checking for pid/kill was introducing some delay which helped in the cleanup of portchannels.
There was a problem hiding this comment.
Tried a 10ms delay, and this seems to work well, at least on Arista 7260cx3.
Signed-off-by: Saikrishna Arcot <[email protected]>
|
/azp run |
|
Azure Pipelines successfully started running 1 pipeline(s). |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What I did
When killing 10 or more teamd processes, add a delay of 0.1 seconds after every 10 kill signals/proceses. This is because in the LAG scale tests (in
ecmp/inner_hashing/test_inner_hashing_lag.pyin sonic-mgmt), it may create 100 LAGs, and when destroying them all, some of those LAGs may fail to be properly destroyed, leaving some stale port channels around. This seems to be because the netlink socket buffers on which the teamd processes get notifications become full with events of the other port channels/interfaces going downWhy I did it
As a workaround, add some delays in killing the teamd processes, so that the netlink buffers don't become full, causing messages to get dropped.
This delay was randomly chosen, and it seems to work well with 100 LAGs on a KVM. It can probably made to be a bit more aggressive if needed (i.e. maybe 0.05 seconds every 20 processes).
How I verified it
On a KVM testbed with t0-116 topology with a bit more than 100 LAGs, stop teamd using
sudo systemctl stop teamd, and verify that all of the LAGs were deleted, and there were no messages from the kernel similar to the following:Details if related
Partial fix for sonic-net/sonic-buildimage#19310.