Hi cmvs,
Thanks for testing NC.
On 2014-03-02 14:39, cmsv wrote:
I just found out /(the hard way) that using network coding (batman-adv 2013.4) that i have a massive, unbelievable bandwidth loss if i have more than one hop to the destination
I tested this in a real scenario with 11 nodes as well as 3 nodes in a house with 3 floors and having one per each floor.
Here is the example with 3 nodes
A <------ B -----> C
From node B to iperf reports me 15 mbit From node B to node C iperf reports me 9 mbit
Then, from node C to node A i get values bellow 1 mbit such as 647kbits or less. More than 3 nodes is to forget.
uci set batman-adv.bat0.network_coding=0 disables nc and solved the problem which brought the network bandwidth to a more realistic and mathematical acceptable bandwidth level.
Is this finding known by others ?
Have you verified that both node A and C have working promiscuous mode? One way to do this is to transmit udp packets from B to A with iperf and count the number of received packets on C with tcpdump.
In addition, you should note that NC doesn't bring a gain for unidirectional flows, so even though TCP sends ACKs in the reverse direction, the gain is probably eaten up by the delay added by node B, in order to increase the chance of coding opportunities.
When you have verified the working promisc mode, can you repeat the test and tell us the results?
And also test with two TCP flows: A -> C and C -> A ?
Thanks!