Re: How to avoid kernel's receive buff full , then UDP packet dropped by kernel?

[Ian Swett <[e-mail]>]

I would advise increasing the value from the default, which is commonly
16k.  1MB seems like a good value, experimentally.

On Fri, Nov 29, 2019 at 6:52 PM Ryan Hamilton <rch=
[e-mail]> wrote:

> On the server side of a QUIC connection, the UDP socket (and hence the
> kernel receive buffer) is shared across multiple connections. As such it'd
> be incredibly complicated to come up with a system to handle this and would
> likely result in unnecessarily throttling throughput. On the client side
> (where there is once QUIC connection per socket) this is unlikely to happen
> if the receive buffer is larger than the BDP of the link. Of course, in
> reality, if the application is reading from the socket frequently enough,
> this is likely a non-issue.
>
> On Fri, Nov 29, 2019 at 4:53 AM Jiuhai Zhang <[e-mail]>
> wrote:
>
>> Is there any good idea to solve this? How can we do flow control like TCP
>> receive window size? User space can not get how many receive buff is used
>>
>