Re: [9front] iounit: bump it across the board

[hiro <23hiro@gmail.com>]

even if they don't share a 9p channel they can block other
communication, if we end up bandwidth limited.

normally up to limits of the tcp fairness mechanism.

On 6/9/22, ori@eigenstate.org <ori@eigenstate.org> wrote:
> Quoth ori@eigenstate.org:
>> Quoth noam@pixelhero.dev:
>> > Quoth ori@eigenstate.org:
>> > > This patch bumps up the iounit in the kernel to 32k, and
>> > > introduces the IOUNIT constant to take advantage of it.
>> >
>> > Any reason I shouldn't bump it to 1M or something locally?
>> > (in the kernel and, in userspace, in non-libthread applications)
>> >
>> > - Noam Preil
>> >
>>
>> Try it out, but keep in mind that on a slower connection,
>> you have to wait for the entire iounit to get sent before
>> you can send the next request -- so, eg, tethering over
>> wifi means that your file IO needs to go through fully
>> before your next mouse message can get sent.
>>
>
> phrased poorly -- but, if your mouse and file system share
> a 9p channel (ie, exportfs), then large messages will block
> all other communication.
>
> the other reason is that you can have a lot of in flight
> 9p buffers at once; up to 64k per connection. Most of the
> time there will be fewer, but not always. But even a few
> hundred 1 meg buffers is a lot of memory on smaller modern
> systems.
>
>