3

First of all, I am fairly new to TCP, sorry if I miss anything obvious. Also, any suggestion to help me investigate are welcome.

from my understanding, if I have a setup:

SENDER------SWITCH------RECEIVER
        A           B

if both link A & B have 0% loss rate & 100Mbps bandwidth but:

scenario 1) A has 1~301ms delay and B has 1ms scenario 1) A has 1~301ms delay and B has 1ms

scenario 2) A has 1~151ms delay and B has 1~151ms scenario 2) A has 1~151ms delay and B has 1~151ms

since the total delay of the network should be the same in both scenarios, shouldn't the graph of throughput against time on sender side looks exactly the same?

take the 60ms line in both graphs for example. Although the total delay in scenario 2 is still 60ms in the network, it never encounters another congestion event between around second 1 and until it hits bandwidth cap, not like scenario 1 which had many.

the two major concern are:

  • why does the shape of low latency line during the early stage be different?
  • why do some previously unable-to-reach-cap cases could now reach the cap?

some background info that might/might not help:

  • I used mininet(virtual network with ubuntu) for the test setup.
  • linux, so TCP CUBIC for congestion avoidance algo.
  • the iperf command used

recvr.cmd('iperf -s -p', port, '> %s/iperf_server.txt' % args.dir, '&') sender.sendCmd('iperf -c %s -p %s -t %d -i 1 -yc > %s/iperf_%s.txt' % (recvr.IP(), port, seconds, args.dir, 'h1'))

  • cwnd for link A 121ms delay, link B 1ms (scenario 1, 120ms line) enter image description here
  • cwnd for link A 61ms delay, link B 61ms (scenario 2, 120ms line) enter image description here
0

I think I might have found the reason, but I am not sure. Any feedback will be welcome.

why does some previously unable-to-reach-cap cases could now reach the cap?

The problem is the switch's buffer size.

  • In scenario 1 (link A has 2X delay), the real bottleneck situation is 2X delay with K buffer size.
  • In scenario 2 (link A&b both have X delay), the real bottleneck situation is X delay with K buffer size.

Since this are lossless links, once I crank the buffer size up every case seems to be able to reach network ceiling.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.