>Hi, > >I am doing experiments about packet classification algorithm and I found I >always get RX Error when the throughput is too high, so I did the following >experiments. > >There are two PC servers (A and B), each of them has a Intel 82599ES 10G >with two ports(1 and 2). And they are connected to each other. I simply run >pktgen on both server. > >When I start one of the server and let it generate 10Gb/s traffic each >port, which is start A1 and A2. I can receive the 10Gb/s traffic on the >other server's every port. > >When I start one port of the two servers and let them generate 10Gb/s >traffic to each other, which is start A1 and B1. Both the port show that >they can receive 10Gb/s and send 10Gb/s traffic. > >But When I start both port on both server, which is start A1, A2, B1, B2, >each port shows it can generate 10Gb/s but it can only receive 6.7Gb/s >traffic and the left 3.3Gb/s are considered RX Error. > >When I stop one of the ports, which is start A1, B1, B2, On server A I >receive 8.1Gb/s on A1, 8.4Gb/s on A2, while A1 is sending 10Gb/s traffic, >the traffic left is considered RX Error. I receive 10Gb/s traffic on B1, >sending 10Gb/s traffic on B1 and B2. > >My parameter is -c 0xff -n 4 -- -P -m "[1:2].0, [3:4].1", but it won't >change when I assign more lcore on RX queue.
Are the cores 1, 2, 3, 4 on different physical cores each or do they share a physical core? One of the issues is the PCI bus on some NICs has a bottle neck on the bus and limits the amount of traffic for a given NIC. This could be your problem, but I do not know if these NICs have that problem. > >I'm thinking it maybe is the parameter problem, such as the Hugepage or >others, is there any solution or advices? > Regards, Keith
