flow-level ECMP may not work properly

[wangshuaizs]

Hi,
I tried to used Flow-level ECMP in fat-tree topology, but I found the total throughput of the receiver was not as much as the maximum. I think you use the original source codes about flow-level ECMP of ns-3, and I ran the same topology using TCP in ns-3 and visualized it. Surprisingly, only some core switches were used, so I thought that flow-level ECMP might not work properly and it is the main cause for low throughput. Have you ever noticed this? or I used it wrongly?

thanks.

[bobzhuyb]

I did not notice it. Would you debug it a bit? The code for ECMP is not complicated. Please keep me updated.

[wangshuaizs]

I have got it.
Flow-level ecmp is based on five-tuple, so the flows that have the same hash value will have the same next hop, denoted as SW_A. If SW_A still has more than one next hops, for example, SW_B and SW_C, these flows that have the same hash value will choose the same next hop, SW_B or SW_C.
I have fixed it by dividing tuple value into the switch id.
tupleValue /= (m_ipv4->GetObject<Node>()->GetId() + 1);