I'm doing R&D in SDN. In this case scenario, we use MPLS labeling, and Open vSwitch as software-switches. We are using 2 server nodes with ovs 2.6.0, with kernel modules loaded, and 2 hosts.
They are directly connected through 1 Gigabit Ethernet connections, and there is arround 1 ms of rtt, and in the case of the first packet less than 3 ms (using ping utility). Im using Iperf3 for doing the tests. The first test is the performance reached without using mpls labeling, and the second test is using mpls labeling. The MTU is adjusted to not doing fragmentation, so this is not the problem. I tried adjusting the congestion window and other parameters like TCP algorithm used.
mar jul 4 12:21:09 CEST 2017
Connecting to host 192.168.20.2, port 5201
[ 4] local 192.168.20.1 port 43526 connected to 192.168.20.2 port 5201
[ ID] Interval Transfer Bandwidth Retr Cwnd
[ 4] 0.00-1.00 sec 112 MBytes 943 Mbits/sec 0 450 KBytes
[ 4] 1.00-2.00 sec 112 MBytes 937 Mbits/sec 0 516 KBytes
[ 4] 2.00-3.00 sec 112 MBytes 938 Mbits/sec 0 571 KBytes
[ 4] 3.00-4.00 sec 112 MBytes 937 Mbits/sec 0 625 KBytes
[ 4] 4.00-5.00 sec 112 MBytes 943 Mbits/sec 0 633 KBytes
[ 4] 5.00-6.00 sec 111 MBytes 933 Mbits/sec 0 633 KBytes
[ 4] 6.00-7.00 sec 111 MBytes 933 Mbits/sec 0 664 KBytes
[ 4] 7.00-8.00 sec 112 MBytes 944 Mbits/sec 0 664 KBytes
[ 4] 8.00-9.00 sec 111 MBytes 933 Mbits/sec 0 697 KBytes
[ 4] 9.00-9.16 sec 18.8 MBytes 977 Mbits/sec 0 697 KBytes
- - - - - - - - - - - - - - - - - - - - - - - - -
[ ID] Interval Transfer Bandwidth Retr
[ 4] 0.00-9.16 sec 1.00 GBytes 939 Mbits/sec 0 sender
[ 4] 0.00-9.16 sec 1022 MBytes 935 Mbits/sec receiver
iperf Done.
<----------->
mar jul 4 12:40:10 CEST 2017
Connecting to host 192.168.20.2, port 5201
[ 4] local 192.168.20.1 port 43530 connected to 192.168.20.2 port 5201
[ ID] Interval Transfer Bandwidth Retr Cwnd
[ 4] 0.00-1.00 sec 203 KBytes 1.66 Mbits/sec 57 2.82 KBytes
[ 4] 1.00-2.00 sec 398 KBytes 3.26 Mbits/sec 124 2.82 KBytes
[ 4] 2.00-3.00 sec 400 KBytes 3.28 Mbits/sec 124 2.82 KBytes
[ 4] 3.00-4.00 sec 319 KBytes 2.61 Mbits/sec 124 2.82 KBytes
[ 4] 4.00-5.00 sec 398 KBytes 3.26 Mbits/sec 126 2.82 KBytes
[ 4] 5.00-6.00 sec 395 KBytes 3.24 Mbits/sec 124 2.82 KBytes
[ 4] 6.00-7.00 sec 398 KBytes 3.26 Mbits/sec 126 2.82 KBytes
[ 4] 7.00-8.00 sec 324 KBytes 2.66 Mbits/sec 124 2.82 KBytes
[ 4] 8.00-9.00 sec 398 KBytes 3.26 Mbits/sec 124 2.82 KBytes
[ 4] 9.00-10.00 sec 400 KBytes 3.28 Mbits/sec 126 2.82 KBytes
- - - - - - - - - - - - - - - - - - - - - - - - -
[ ID] Interval Transfer Bandwidth Retr
[ 4] 0.00-10.00 sec 3.55 MBytes 2.98 Mbits/sec 1179 sender
[ 4] 0.00-10.00 sec 3.42 MBytes 2.87 Mbits/sec receiver
I know there are issues using MPLS and using ovs, but there are some facts that are weird in this case:
- If i use UDP instead of TCP, there is one packet out of order, but the rest are good, so packets are using kernel datapath i guess.
- There are 9 packets lost at the start of the TCP transmission, and there are more packets lost periodically. Looking the tcpdump traces, those packets are "missing" in the first node, because in the second hop they are not captured.
- As you can see above, the performance using TCP w/o MPLS labeling is very good.
Is there someone who knows how to tackle this problem?
PD: Excuse me for my possible english typos.