Hide Forgot
Description of problem: rhel6.1 has 2 times higher UDP datagram loss when receiving 1400Bytes long datagrams over 10gig ethernet than rhel6.0 Version-Release number of selected component (if applicable): 2.6.32-131.0.1.el6.x86_64 How reproducible: 100% Steps to Reproduce: 1.run /root/prepare_sys.py 2.run netperf -H 172.16.29.20 -t UDP_STREAM -- -m 1000 Socket Message Elapsed Messages Size Size Time Okay Errors Throughput bytes bytes secs # # 10^6bits/sec 124928 1000 10.00 5026535 0 4021.23 124928 10.00 2139250 1711.40 Actual results: RHEL 6.1 Expected results: datagram loss decreased to rhel6.0 levels Additional info: Please see: http://download.englab.brq.redhat.com/perf-results//netperf/RHEL61-20110413/#u6_fd29--10_to_fd29--20_tp on this plot is shown local send throughput of rhel6.0 and 6.1. Both systems sends almost same amount of data. On this plot http://download.englab.brq.redhat.com/perf-results//netperf/RHEL61-20110413/#u6_fd29--10_to_fd29--20_rtp received traffic is shown. Rhel6.1 receives significantly smaller amount of data. On this plot http://download.englab.brq.redhat.com/perf-results//netperf/RHEL61-20110413/#u6_fd29--10_to_fd29--20_lt is shown how much network traffic is lost. It is clear that on rhel6.1 amount of lost traffic is much higher.
I presume the 82599 interfaces were the only ones that saw this drop? I would be curious to know how the RHEL6.0 throughput compares to 2.6.32-71.27.1.el6 or later. There are changes in RHEL6.1 and 2.6.32-71.27.1.el6 that fix an erratum, but likely impact performance. Would you be able to run a similar test with RHEL6.0.z kernel 2.6.32-71.27.1.el6 or later to help verify?