How to set udp buffer size for a docker container?

Question

I would like to set the UDP buffer size for a docker container. Though, the UDP buffer size is a setting related to the kernel and it corresponds to that of the host, I see that, I don't see my host setting of the buffer size affecting my container.

Here is what I have tried:

On host

sysctl -w net.core.rmem_default = 4194304
sysctl -w net.core.rmem_max = 4194304

and to verify (on host)

sysctl net.core.rmem_default
net.core.rmem_default = 4194304

sysctl net.core.rmem_max
net.core.rmem_max = 4194304

But, when I run in the container (say alpine).

docker run -it alpine

and in alpine ash

sysctl net.core.rmem_default
sysctl: error: 'net.core.rmem_default' is an unknown key

The same is the case with rmem_max also.

However, to my surprise, I see

sysctl net.core.rmem_default
net.core.somaxconn = 128

I am able to set the above on docker container using --sysctl argument.

docker run -it --sysctl net.core.somaxconn=125 alpine
/ # sysctl net.core.somaxconn
net.core.somaxconn = 125

But, what I need is rmem_default and rmem_max sizes.

If this option is not provided through --sysctl flag and that the setting depends on the host machine, then

How can I prove that the buffer size on the container is same as that of the host?

Moreover, during many tests, I came to know that the packets are being dropped (I think, it is because of some low buffer size being set).

I also tried writing to the /proc/sys/net/core/rmem_default in the container but I got cannot stat error. I did not understand why there is no such file as rmem_default and the like (rmem_max etc) which are found on the host in the docker /proc.

I have heard that the host /proc is mounted on the container. Is it true? It doesn't seem to be considering the above scenario.

Also, editing the sysctl.conf and calling sysctl -p also did not help.

Feel free to move this question to unix.stackexchange.com if it makes sense there.

Answer 1

Your main problem is that those values are not namespaced, that means they are shared between all network namespaces. There is a kernel patch in state RFC since 2017 which wants to make them namespaced: [RFC] net: make net.core.{r,w}mem_{default,max} namespaced

This patch alone would not allow you to use them within a kubernetes pod. You would need to allow the sysctls like described in Using sysctls in a Kubernetes Cluster