[SOLVED] another tc question, this time about behavior if 2 members of same class are sending

psycroptic · 09-17-2013, 04:57 PM

so this is what's going on:

Code:

tc qdisc add dev internal root handle 1: htb default 10
tc class add dev internal parent 1: classid 1:1 htb rate 51.88mbit ceil 51.88mbit
tc class add dev internal parent 1:1 classid 1:10 htb rate 34.38mbit ceil 51.88mbit
tc class add dev internal parent 1:1 classid 1:11 htb rate 12.5mbit ceil 51.88mbit
tc class add dev internal parent 1:1 classid 1:12 htb rate 4.4mbit ceil 51.88mbit

say that 2 members of the default class (1:10) are downloading, and both download streams are capable of >50mbps. I would think that, since the whole class is guaranteed 34.38mbit, each client would balance out at 34.38/2=~17.19mbps.

but that's not what happens; it appears that they each fluctuate between 1 having almost all the ceil bandwidth and the other having a tiny amount. they go back and forth at a somewhat random rate, so the end result is the download speed on each client rises and falls.

why is this? is this to be expected?

baldy3105 · 09-18-2013, 11:43 AM

This is what you expect to see in a fifo queue. TCP responds to latency and packet loss by slowing down. As one slows down the other picks up the slack, speeding up until it bounces off the bandwidth limit thus increasing the latency it experiences and it throttles back again allowing the other flow to come to the fore. The two TCP flows are both attempting to optimise their thoughput and they are not co-ordinated.

Within a class you get fifo queueing, the only prioritisation or bandwidth occurs between queues, not within a queue.

psycroptic · 09-18-2013, 02:43 PM

I see.

So there wouldn't be a way to force them to balance out evenly then?
EDIT: suppose not.

Well thanks in any case, and also for that other post that I think you replied to.

baldy3105 · 09-19-2013, 02:22 PM

This is actually a subject that can get horrifically complex very quickly. The behaviour of speeding up and slowing down in an effort to maximise throughput in TCP is a combination of "Slow start" and "Congestion Avoidance" behaviour, both designed to even out transmission behaviour.

I've seen degree dissertations done by university graduates that are basically mathematical discussions on how the current methods might be improved!

The thing to worry about is the average thoughput over the duration of the flow, rather than the fluctuations, otherwise you'll give yourself a headache!

psycroptic · 09-19-2013, 02:41 PM

word, thanks for the info. yeah, it's kind of been making my head spin trying to think about traffic flows and where in the chain I need to apply various rules. it's been a good learning experience for understanding how networks operate conceptually.

if you've got any ideas on my new project of limiting upload, feel free to chime in