0

I am trying to understand TCP congestion avoidance mechanism, but I don't understand one thing: TCP congestion avoidance is per flow or per link?

In other words: there are 2 routers A and B A is sending B two TCP flows - when one TCP flow detects congestion, does it decrease window size in the other flow as well?

of course if this happens, the other flow will detect congestion in some time, but does the second flow "waits" until it detects congestion on its own? that would be quite uneffective...

thanks a lot

Igor
  • 592
  • 2
  • 9
  • 31

2 Answers2

1

It decreases the window size for the current connection. Each connection's RTT and windows are maintained independently.

user207421
  • 305,947
  • 44
  • 307
  • 483
1

Routers operate on layer 3 (IP) and are not aware of layer 4 (TCP), because of this, routers do not take any part in TCP congestion avoidance mechanism. This mechanism is fully implemented by TCP endpoints. It is triggered by routers dropping IP packets, but (classic) routers are not aware what higher level protocol IP packets carry.

The fact that one flow does not affect the other is quite desirable from the security perspective. With NAT you can have many hosts sharing the same IP address. From the outside world all these hosts look as a single machine. So, if some server reduced throughput of all TCP connections coming from a single IP address in response to packets dropped within one of those connections that would open a door to quite nasty DoS attacks.

Another issue is that some routers may be configured to drop packets based on IP ToS field. For example, latency sensitive SSH traffic may set different ToS than bulk FTP download. If router is configured to take into account ToS field, it may drop packets belonging to FTP connection, which should trigger congestion avoidance, but is should not affect packets belonging to SSH connection, which may be handled with higher priority.

Jan Wrobel
  • 6,969
  • 3
  • 37
  • 53