Extreme slowness in IPIP tunnel when routing through kernel 3.18 and later

[Patrick Boutilier]

I have an IPIP tunnel setup between two hosts in different buildings. 
The Linux router they route through causes extreme slowness in the 
tunnel when running kernels from 3.18 on . tcpdump shows many cksum 
errors which don't show up in the 3.17 and earlier kernels. Speed goes 
back to normal when rx-checksumming and tx-checksumming are turned off 
using ethtool on the Linux router . Would this be an effect of bulk 
network packet transmission that was introduced in 3.18 or some other issue?

eth2
driver: igb
version: 5.2.15-k
firmware-version: 1.5.1
bus-info: 0000:06:00.0
supports-statistics: yes
supports-test: yes
supports-eeprom-access: yes
supports-register-dump: yes
supports-priv-flags: no


eth6
driver: ixgbe
version: 3.19.1-k
firmware-version: 0x80000389
bus-info: 0000:41:00.0
supports-statistics: yes
supports-test: yes
supports-eeprom-access: yes
supports-register-dump: yes
supports-priv-flags: no





Example of cksum from tcpdump:

Flags [.], cksum 0x2237 (incorrect -> 0xbb02), seq 424119200:424120648, 
ack 1, win 115, options [nop,nop,TS val 303173784 ecr 3971495454], 
length 1448





Results of nuttcp test with rx-checksumming and tx-checksumming on:


Testing upload inside tunnel



    0.0625 MB /   5.00 sec =    0.1048 Mbps    79 retrans
    0.0625 MB /   5.00 sec =    0.1048 Mbps    59 retrans
    0.0625 MB /   5.00 sec =    0.1049 Mbps    27 retrans

    0.2096 MB /  16.11 sec =    0.1092 Mbps 0 %TX 0 %RX 186 retrans 
0.47 msRTT



Testing download inside tunnel



    0.5000 MB /   5.00 sec =    0.8389 Mbps   112 retrans
    2.6250 MB /   5.00 sec =    4.4040 Mbps   412 retrans
    0.8750 MB /   5.00 sec =    1.4680 Mbps   156 retrans

    4.0243 MB /  15.48 sec =    2.1809 Mbps 0 %TX 0 %RX 705 retrans 
0.47 msRTT


Testing upload outside tunnel



  423.0000 MB /   5.00 sec =  709.4949 Mbps   275 retrans
  442.0625 MB /   5.00 sec =  741.7008 Mbps    53 retrans
  418.5625 MB /   5.00 sec =  702.1329 Mbps   133 retrans

 1287.0080 MB /  15.08 sec =  715.9252 Mbps 1 %TX 30 %RX 461 retrans 
0.43 msRTT



Testing download outside tunnel



  552.4375 MB /   5.00 sec =  926.8347 Mbps    26 retrans
  555.1875 MB /   5.00 sec =  931.4499 Mbps    17 retrans
  553.0000 MB /   5.00 sec =  927.7802 Mbps    16 retrans

 1664.2149 MB /  15.03 sec =  928.6737 Mbps 9 %TX 5 %RX 60 retrans 0.46 
msRTT






Results of nuttcp test with rx-checksumming and tx-checksumming off:



Testing upload inside tunnel



  440.0000 MB /   5.00 sec =  738.0112 Mbps   117 retrans
  433.0000 MB /   5.00 sec =  726.4892 Mbps   135 retrans
  447.9375 MB /   5.00 sec =  751.5714 Mbps   112 retrans

 1321.0069 MB /  15.04 sec =  736.7213 Mbps 3 %TX 21 %RX 364 retrans 
0.51 msRTT



Testing download inside tunnel



  494.0625 MB /   5.00 sec =  828.8973 Mbps   335 retrans
  471.0000 MB /   5.00 sec =  790.2050 Mbps   219 retrans
  469.8750 MB /   5.00 sec =  788.3174 Mbps   220 retrans

 1435.1875 MB /  15.00 sec =  802.4203 Mbps 89 %TX 12 %RX 774 retrans 
0.46 msRTT


Testing upload outside tunnel



  431.2500 MB /   5.00 sec =  723.2808 Mbps    33 retrans
  446.3750 MB /   5.00 sec =  748.9268 Mbps    10 retrans
  475.4375 MB /   5.00 sec =  797.7044 Mbps     0 retrans

 1355.0834 MB /  15.06 sec =  754.6438 Mbps 1 %TX 33 %RX 43 retrans 
0.53 msRTT



Testing download outside tunnel



  424.8125 MB /   5.00 sec =  712.7061 Mbps   102 retrans
  456.5625 MB /   5.00 sec =  765.9872 Mbps    80 retrans
  454.0000 MB /   5.00 sec =  761.6935 Mbps   116 retrans

 1335.7823 MB /  15.01 sec =  746.6959 Mbps 4 %TX 5 %RX 298 retrans 
0.55 msRTT
<<attachment: boutilpj.vcf>>