Fast IP Forwarding to WAN, but falls off cliff between LAN subnets

Question

I replaced my consumer wireless router with a linux box that has a quad-gigabit NIC PCIe card and a single gigabit NIC on the motherboard (for the WAN). After turning on IP forwarding, masquerading (via iptables), and setting up subnets on each of the four LAN interfaces I ran some speed tests.

$ ip route
default dev ppp0 scope link 
10.0.0.0/16 dev enp3s0f0 proto kernel scope link src 10.0.0.1 
10.64.0.0/16 dev enp3s0f1 proto kernel scope link src 10.64.0.1 
10.192.0.0/16 dev enp4s0f1 proto kernel scope link src 10.192.0.1 
aaa.bbb.ccc.ddd dev ppp0 proto kernel scope link src www.xxx.yyy.zzz 

• From a wireless device on one of the LAN subnets to a speedtest server on the WAN I get the full 40 Mbps / 5 Mbps I pay my ISP for.

• From the router host to a wired LAN host using iperf3 I can consistently maintain 930+ Mbps for several minutes.

• From a wired device on one of the LAN subnets to a wired device on a different LAN subnet using iperf3 I initially get 80-95 Mbps for the first few seconds but it rapidly drops to zero.

• From a wired device on one of the LAN subnets to a wired device on a different LAN subnet using iperf3 with a target bitrate of 20 Mbps I see the similar results (see update at end), but can sustain about 10 Mpbs

.

Connecting to host 10.0.0.2, port 5201
[  5] local 10.192.128.3 port 35620 connected to 10.0.0.2 port 5201
[ ID] Interval           Transfer     Bitrate         Retr  Cwnd
[  5]   0.00-1.00   sec  10.2 MBytes  85.9 Mbits/sec    0   73.5 KBytes       
[  5]   1.00-2.00   sec  9.01 MBytes  75.6 Mbits/sec    0   82.0 KBytes       
[  5]   2.00-3.00   sec  8.26 MBytes  69.3 Mbits/sec    0   79.2 KBytes       
[  5]   3.00-4.00   sec  9.01 MBytes  75.6 Mbits/sec    0   73.5 KBytes       
[  5]   4.00-5.00   sec  5.28 MBytes  44.3 Mbits/sec    1   1.41 KBytes       
[  5]   5.00-6.00   sec  0.00 Bytes  0.00 bits/sec    1   1.41 KBytes       
[  5]   6.00-7.00   sec  0.00 Bytes  0.00 bits/sec    1   1.41 KBytes       
[  5]   7.00-8.00   sec  0.00 Bytes  0.00 bits/sec    1   1.41 KBytes       
[  5]   8.00-9.00   sec  0.00 Bytes  0.00 bits/sec    0   1.41 KBytes       
^C[  5]  10.00-13.63  sec  0.00 Bytes  0.00 bits/sec    1   1.41 KBytes       
- - - - - - - - - - - - - - - - - - - - - - - - -
[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-13.63  sec  41.8 MBytes  25.7 Mbits/sec    5             sender
[  5]   0.00-13.63  sec  0.00 Bytes  0.00 bits/sec                  receiver
iperf3: interrupt - the client has terminated

This is suggesting to me that there's some problem forwarding packets between the subnets. I first ensured that my iptables rules are as minimal as possible:

-t nat -A POSTROUTING -o ppp0 -j MASQUERADE
# WAN connection is PPPoE and VLAN tagged
-t filter -A FORWARD -o ppp0 -p tcp --tcp-flags SYN,RST SYN -j TCPMSS  --clamp-mss-to-pmtu

Dumping the iptables state I see low packet counts for both rules.

Next I checked for packet loss. There does seem to be a small but consistent amount of packet loss / retransmits.

$ sudo netstat -s | egrep -i 'retransmit|drop'
    498 outgoing packets dropped
    25848 fast retransmits

I then thought that maybe there was a buffer or queue that was filling and packets were getting dropped. I calculated the average bandwidth-delay product and compared that against the reserved memory.

$ sudo ping -f 10.0.0.2 -s $((1500-28))               
PING 10.0.0.2 (10.0.0.2) 1472(1500) bytes of data.
.^C
--- 10.0.0.2 ping statistics ---
9036 packets transmitted, 9035 received, 0% packet loss, time 26512ms
rtt min/avg/max/mdev = 1.742/2.817/12.057/0.758 ms, pipe 2, ipg/ewma 2.934/3.091 ms

$ echo "1*(1024^3) * 0.003" | bc 
3221225.472

$ cat /proc/sys/net/ipv4/tcp_mem
18396   24529   36792

$ getconf PAGESIZE
4096

That appears to be sufficient. So now I'm a bit stuck. I ran tcpdump on the iperf3 client and can see things moving along well for a bit. Then I see a long (almost 250ms) period of silence before lots of retransmits and duplicate acknowledgements.

Since I can pull sufficient download speeds from the WAN I don't suspect that the onboard NIC is at fault. I'm looking for help to diagnose this quad-NIC (details below) and possibly a dumb layer-2 gigabit switch (Netgear GS-108) and any other kernel configuration that could be getting in the way. I doubt it's the switch, as it's never been a problem before and I can maintain speeds from the router's loopback to that subnet. Only inter-subnet performance appears to be affected.

  *-network:0               
       description: Ethernet interface
       product: 82571EB Gigabit Ethernet Controller (Copper)
       vendor: Intel Corporation
       physical id: 0
       bus info: pci@0000:03:00.0
       logical name: enp3s0f0
       version: 06
       serial: 00:26:55:xx:xx:xx
       size: 1Gbit/s
       capacity: 1Gbit/s
       width: 32 bits
       clock: 33MHz
       capabilities: pm msi pciexpress bus_master cap_list ethernet physical tp 10bt 10bt-fd 100bt 100bt-fd 1000bt-fd autonegotiation
       configuration: autonegotiation=on broadcast=yes driver=e1000e driverversion=3.2.6-k duplex=full firmware=5.12-2 ip=10.0.0.1 latency=0 link=yes multicast=yes port=twisted pair speed=1Gbit/s
       resources: irq:24 memory:fe920000-fe93ffff memory:fe880000-fe8fffff ioport:d020(size=32)

UPDATE:

$ iperf3 -b 20m -c 10.0.0.2
Connecting to host 10.0.0.2, port 5201
[  5] local 10.192.128.3 port 36554 connected to 10.0.0.2 port 5201
[ ID] Interval           Transfer     Bitrate         Retr  Cwnd
[  5]   0.00-1.00   sec  2.49 MBytes  20.9 Mbits/sec    0    158 KBytes       
[  5]   1.00-2.00   sec  2.38 MBytes  19.9 Mbits/sec    0    150 KBytes       
[  5]   2.00-3.00   sec  2.38 MBytes  19.9 Mbits/sec    1    133 KBytes       
[  5]   3.00-4.00   sec  2.38 MBytes  19.9 Mbits/sec    0   73.5 KBytes       
[  5]   4.00-5.00   sec  2.38 MBytes  19.9 Mbits/sec    0   70.7 KBytes       
[  5]   5.00-6.00   sec  1.12 MBytes  9.44 Mbits/sec    2   1.41 KBytes       
[  5]   6.00-7.00   sec  0.00 Bytes  0.00 bits/sec    2   1.41 KBytes       
[  5]   7.00-8.00   sec  0.00 Bytes  0.00 bits/sec    0   1.41 KBytes       
[  5]   8.00-9.00   sec  0.00 Bytes  0.00 bits/sec    1   1.41 KBytes       
iperf3: error - control socket has closed unexpectedly

$ iperf3 -b 10m -c 10.0.0.2 
Connecting to host 10.0.0.2, port 5201
[  5] local 10.192.128.3 port 36564 connected to 10.0.0.2 port 5201
[ ID] Interval           Transfer     Bitrate         Retr  Cwnd
[  5]   0.00-1.00   sec  1.24 MBytes  10.4 Mbits/sec    0    201 KBytes       
[  5]   1.00-2.00   sec  1.25 MBytes  10.5 Mbits/sec    0    118 KBytes       
[  5]   2.00-3.00   sec  1.12 MBytes  9.44 Mbits/sec    0    127 KBytes       
[  5]   3.00-4.00   sec  1.25 MBytes  10.5 Mbits/sec    0    107 KBytes       
[  5]   4.00-5.00   sec  1.12 MBytes  9.44 Mbits/sec    0    110 KBytes       
[  5]   5.00-6.00   sec  1.25 MBytes  10.5 Mbits/sec    0   90.0 KBytes       
[  5]   6.00-7.00   sec  1.12 MBytes  9.44 Mbits/sec    0   87.2 KBytes       
[  5]   7.00-8.00   sec  1.25 MBytes  10.5 Mbits/sec    0   81.6 KBytes       
[  5]   8.00-9.00   sec  1.12 MBytes  9.44 Mbits/sec    0   78.8 KBytes       
[  5]   9.00-10.00  sec  1.25 MBytes  10.5 Mbits/sec    0    112 KBytes       
- - - - - - - - - - - - - - - - - - - - - - - - -
[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-10.00  sec  12.0 MBytes  10.1 Mbits/sec    0             sender
[  5]   0.00-10.04  sec  12.0 MBytes  10.0 Mbits/sec                  receiver

iperf Done.

Answer 1

Thanks to @Pedro for helping me dig in. Originally I thought this was a bad piece of hardware, but after replacing it with another I'm certain it's a driver problem. I'm still digging in to find out if this is a bug that's been reported already or not (and whether a fix exists). In the mean time, I did locate a serverfault question which linked to a bug report that suggested turning off several offloading features. This at least got me from 0 bps to ~270 Mbps stably. Far short of the ~940 Mbps it is capable of, but better than nothing while I continue researching.

ethtool -K eth0 gso off gro off tso off