UDP receive buffer error spikes

[Marcin Owsiany]

Hello,

I noticed recently (credit goes to netdata) that during high chrome activity the kernel reports lots of UDP read buffer errors. It seems like this tends to happen when there is a high CPU usage combined with relatively high incoming UDP packet rate.

The following example is from when I restarted chrome after an upgrade. The CPU was nearly maxed out while chrome was starting up. Note how ~400 incoming pps translates to about twice as many UDP receive buffer errors per second! At the same time, netdata reported that the amount of memory used by allocated UDP sockets (ipv4.sockstat_udp_mem) was quite low (hovering around 2MB).

This on a rather old laptop (dual-core Intel(R) Core(TM) i3-3110M CPU @ 2.40GHz but with 16 GB of RAM and an SSD) running Debian 12. It was running with the default net.core.rmem_max=212992. I tried reproducing the issue by navigating to "about:restart" with the sysctl value bumped 10x but the packet rate was much lower, so it did not trigger the issue.

I wonder whether others have seen issues like this. I did not see any user-visible symptoms, but perhaps this is something that can be improved to avoid the high packet drop rate and retransmissions...

regards,

Marcin

[Marcin Owsiany]

wtorek, 5 września 2023 o 19:12:16 UTC+2 Marcin Owsiany napisał(a):

Note how ~400 incoming pps translates to about twice as many UDP receive buffer errors per second!

I confused the graphs, that should have been ~800 dropped pps out of ~2500 incoming pps, but the the question still stands: doesn't chromium need to improve its UDP flow control?

Marcin