[LibreQoS] [Starlink] [Rpm] [EXTERNAL] Re: Researchers Seeking Probe Volunteers in USA

[Sebastian Moeller]

Hi Bob,


> On Mar 13, 2023, at 20:32, rjmcmahon <rjmcmahon at rjmcmahon.com> wrote:
> 
> On 2023-03-13 11:51, Sebastian Moeller wrote:
>> Hi Bob,
>>> On Mar 13, 2023, at 19:42, rjmcmahon <rjmcmahon at rjmcmahon.com> wrote:
>>>> 	[SM] not really, given enough capacity, typical streaming protocols
>>>> will actually not hit the ceiling, at least the one's I look at every
>>>> now and then tend to stay well below actual capacity of the link.
>>> I think DASH type protocol will hit link peaks. An example with iperf 2's burst option a controlled WiFi test rig, server side first.
>> 	[SM] I think that depends, each segment has only a finite length and
>> if this can delivered before slow start ends that burst might never
>> hit the capacity?
>> Regards
> 
> I believe most CDNs are setting the initial CWND so TCP can bypass slow start.

	[SM] My take is not necessarily to bypass slow start, but to kick it off with a higher starting point... which is the conservative way to expedite slow-start. Real man actually increase the multiplication factor instead, but there are few real men around (luckily)... s I see both the desire to finish many smaller transfers within the initial window (so the first RTT after the handshake IIUC).

> Slow start seems an engineering flaw from the perspective of low latency.

	[SM] Yes, exponential search, doubling every RTT is pretty aggressive.


> It's done for "fairness" whatever that means.

	[SM] It is doe because:
a) TCP needs some capacity estimate
b) preferably quickly
c) in a way gentler than what was used before the congestion collapse.
	we are calling it slow-start not because it is slow in absolute terms (it is pretty aggressive already)
	but IIUC because before slow start people where even more aggressive (immediately sending at line rate?)

I think we need immediate queue build-up feedback so each flow can look at its own growth projection and the queue space shrinkage projection and then determine where these two will meet. Essentially we need  a gently way of ending slow-start instead of the old chestnut, dump twice as much data in flight into the network before we notice... it is this part that is at odds with low latency. L4s, true to form with its essential bit-banging of queue filling status over all flows in the LL queue is essentially givvin too little information too late. 


If I had a better proposal for a slow-start altenative I would make it, but for me slow-start is similar to what Churchill is supposed to have said about democracy "democracy is the worst form of government – except for all the others that have been tried."...


> 
> Bob