We have a device outputting high speed data over UDP connected directly to a processor (no switch). The processor is running CentOS 6.6. The connection is gigabit Ethernet. When we use packet sizes in the 1400 byte range, everything works great. When we switch to jumbo packets - packet size just under 9000 bytes - we sometimes have receive errors. By sometimes, I mean that if there are no receive errors at startup, they do not eventually start occurring but if there are receive errors at startup, they do not stop. When we have receive errors, only some of the data gets through. Running ethtool -S on the interface results in the following non-zero values:
rx_packets: 4752316
tx_packets: 7407
rx_bytes: 42021872746
tx_bytes: 477144
rx_broadcast: 34
tx_broadcast: 8
tx_multicast: 37
rx_errors: 6161970
rx_crc_errors: 3080985
rx_csum_offload_good: 7325
rx_header_split: 4752250
We have tried swapping out devices and cables with no change. We have also upgraded to the latest e1000e driver. The same scenario works fine on one of our test machines that is also running CentOS 6.6. Both systems are using the Intel 82574L Ethernet controller. Does anyone have any suggestions on how to fix this problem?
