[Make-wifi-fast] 3g/4g with fq_codel & mobility

[Make-wifi-fast] 3g/4g with fq_codel & mobility

[Loganaden Velvindron]

Hi folks,

I have one of those gl.inet 6416a board. It runs fq_codel on the
USB-3g dongle modem interface, and it runs very well at rest.

I put it inside a car, and while driving, I notice some induced latency:

https://www.dslreports.com/speedtest/6354936

Can someone who has some understanding of 3g/4g internals shed some
light ? Could the 3g/4g modem be buffering due to varying strength of
the signal or due to handover in base station ?

Re: [Make-wifi-fast] 3g/4g with fq_codel & mobility

[Jonathan Morton]

> On 21 Nov, 2016, at 13:16, Loganaden Velvindron <loganaden@gmail.com> wrote:
> 
> I have one of those gl.inet 6416a board. It runs fq_codel on the
> USB-3g dongle modem interface, and it runs very well at rest.
> 
> I put it inside a car, and while driving, I notice some induced latency:
> 
> https://www.dslreports.com/speedtest/6354936
> 
> Can someone who has some understanding of 3g/4g internals shed some
> light ? Could the 3g/4g modem be buffering due to varying strength of
> the signal or due to handover in base station ?

Usually such dongles don’t give direct access to the 3G modem interface.  Instead they provide an Ethernet or PPP interface, and do a conversion internally.  Ethernet emulation is more common of late, possibly due to the rise of 4G.

Running fq_codel on such a device without a shaper relies on the buffers in *both* the dongle itself and the driver for the emulated Ethernet device having appropriately sized buffers in aggregate.  This is virtually never the case.  And that’s just for the upload direction.

In the download direction you are relying on buffers in the provider’s network and the cell tower.  Fq_codel on your own hardware has no effect here.

To solve these problems you have to take control of the bottleneck, ie. to move it to a location you control using a pair of shapers (one in each direction).  That’s what I do, using Cake of course.  The difficulty is then in determining the link speed to shape at, as that will change based on propagation conditions, interference, distance, and antenna orientations.

It’s a difficult problem, and one that is so far inadequately solved.

 - Jonathan Morton

Re: [Make-wifi-fast] 3g/4g with fq_codel & mobility

[Loganaden Velvindron]

On Mon, Nov 21, 2016 at 3:27 PM, Jonathan Morton <chromatix99@gmail.com> wrote:
>
>> On 21 Nov, 2016, at 13:16, Loganaden Velvindron <loganaden@gmail.com> wrote:
>>
>> I have one of those gl.inet 6416a board. It runs fq_codel on the
>> USB-3g dongle modem interface, and it runs very well at rest.
>>
>> I put it inside a car, and while driving, I notice some induced latency:
>>
>> https://www.dslreports.com/speedtest/6354936
>>
>> Can someone who has some understanding of 3g/4g internals shed some
>> light ? Could the 3g/4g modem be buffering due to varying strength of
>> the signal or due to handover in base station ?
>
> Usually such dongles don’t give direct access to the 3G modem interface.  Instead they provide an Ethernet or PPP interface, and do a conversion internally.  Ethernet emulation is more common of late, possibly due to the rise of 4G.
>
> Running fq_codel on such a device without a shaper relies on the buffers in *both* the dongle itself and the driver for the emulated Ethernet device having appropriately sized buffers in aggregate.  This is virtually never the case.  And that’s just for the upload direction.
>
> In the download direction you are relying on buffers in the provider’s network and the cell tower.  Fq_codel on your own hardware has no effect here.
>

I think that reducing the buffer size with NCM might be a good start,
particularly with 4G. THere is a lot of work to be done on the
USB-ethernet stack.