Current uses of QUIC benefit from higher buffer size limits. Please
consider setting:

net.core.rmem_max = 7500000
net.core.wmem_max = 7500000

See https://github.com/quic-go/quic-go/wiki/UDP-Buffer-Sizes

Control: tag -1 moreinfo

I can see that it makes sense to increase this to allow increasing the
receive window if the QUIC implementation sets its receive window based
on this, or otherwise to reduce the risk of packets being dropped if
user-space is slow to handle them.

I would like to see a detailed justification for the specific value,
though.

This seems wrong.  With QUIC the memory for replaying lost packets is
allocated in user-space, so I don't see any need to increase buffering
on the send side.  I think that this would mostly contribute to
bufferbloat.

No justification is given there for increasing wmem_max.

Ben.

clone 1111052 -1
retitle 1111052 Please increase default UDP receive buffer size (rmem_max)
tags 1111052 - moreinfo
retitle -1 Please increase default UDP send buffer size (wmem_max)
thanks

I haven't seen a specific justification for that *exact* value, other
than that popular userspace tools/libraries seem to request 7MiB (note:
not 7MB) by default. I expect that 7340032 then got rounded up to
7500000.
and the arguments made there only support a change to rmem_max.

I've split this issue into two: one for rmem_max which has justification
provided, and one for wmem_max which I've left the moreinfo tag on.

Upstream, in
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=a6d4f25888b83
, changed these both to 4MiB (4194304).

Please consider backporting this change.

Thanks,
Josh Triplett

Hi Josh,

We usually wont't take patches from mainline if they are not going to
land in stable series. I think the best course of action if you want
to to see that in a stable series corresponding version is to argue
with upstream why it should be backported upstream to a stable series
as well.

If you do so, can you include us please in CC so we are aware of the
progress?

Regards,
Salvatore

As this is now in the kernel in unstable, do you still also want it in
linux-base?  (Since it might be useful for those running an older vendor
or custom kernel.)

Ben.