abrupt speed changes and TCP

Thu Jan 30 20:01:43 UTC 2020

On 1/30/20 11:46 AM, William Herrin wrote:
> On Thu, Jan 30, 2020 at 10:58 AM Michael Thomas <mike at mtcc.com> wrote:
>> So it occurs to me in the rollout of 5G just walking down the street you
>> might shift back and forth between high speed 5G bands and 4G because of
>> uneven deployment and all sorts of other reasons. It sounds like this
>> could vary block by block practically.
>>
>> I assume TCP just views this as congestion? But with all of the
>> congestion avoidance algorithms and the rapidly fluctuating bandwidth,
>> wouldn't that result in the sender essentially adapting to the least
>> common denominator (eg 4G)? The same goes with latency, I suppose for
>> real time apps.
> Hi Mike,
>
> TCP speed is about two things: round trip time and packet loss.
>
> If the round trip time gets long and then gets short again, TCP will
> immediately adjust. It doesn't much care about the clock time, it
> cares about whether it has received the ack it was looking for.
>
> When packets are lost... trouble. TCP starts with 10 packets, waits
> for an ack, doubles to 20 packets, waits for an ack, doubles to 40
> packets, etc. This is called the congestion window and that early
> phase is called "slow start." When the first packet is lost, that
> doubling growth stops hard. Slows down to one or two additional
> packets per round trip time. Then there's later packet loss.. get very
> much and the congestion window starts halving and only growing pack at
> a packet or two per round trip time. So if you drop packets switching
> back and forth between 4G and 5G, that TCP connection will slow to a
> crawl.
>
> And just for fun: packet reordering is often interpreted as packet
> loss. So if your 5G packet beats your earlier 4G packet to the
> destination, it's as if you lost a packet.
>
>

I kinda hope that Fred will chime in here because he's usually clued in
about research stuff. I suppose at some level we've been here with the
transition between 3 and 4G, but streaming, etc were a lot less common
then. So do you think it would effectively settle into the least common
denominator or would it be an ongoing cluster fuck?

Mike