Uploaded image for project: 'OpenShift Bugs'
  1. OpenShift Bugs
  2. OCPBUGS-30925

Higher than expected TCP Retransmissions with multithreaded loads

XMLWordPrintable

    • No
    • False
    • Hide

      None

      Show
      None

      Description of problem:

          When a pod network load is applied with parallel connections (not just a single connection), running at 4096 byte message sizes.

      Version-Release number of selected component (if applicable):

          MicroShift 4.15.0-rc5

      How reproducible:

          Consistently

      Steps to Reproduce:

          1. Set up MicroShift node (`dnf install microshift`)
          2. Start MicroShift service (`systemctl start microshift`)
          3. Run super-netperf or similar to test pod to pod network:
      super-netperf 2 -H <pod_ip> -l 30 -t TCP_STREAM -- -k rt_latency,p99_latency,throughput,throughput_units,remote_recv_calls,local_send_calls,local_transport_retrans -m 4096 -R 1    

      Actual results:

      We can see relatively large retransmit values, in this example below 86.8
      
          $ cat loss-rt-result-1710248752.csv
      Type,Driver,Profile,Same node,Host Network,Service,Duration,Parallelism,# of Samples,Message Size,Confidence metric - low,Confidence metric - high,Value
      TCP Retransmissions,netperf,TCP_STREAM,true,false,false,30,2,5,4096,5712.581064693677,5892.130935306324,86.800000

      Expected results:

      We do not expect to see high retransmits.

      Additional info:

          We can also use the network-perf v2 workload:
      https://github.com/cloud-bulldozer/e2e-benchmarking/tree/master/workloads/network-perf-v2
      
      With this configuration:
      $ cat one.yaml
      ---
      tests:
        - TCPStreamMed2:
          parallelism: 2
          profile: "TCP_STREAM"
          duration: 30
          samples: 5
          messagesize: 4096

       

            sdn-team-bot sdn-team bot
            sejug2 Sebastian Jug
            Anurag Saxena Anurag Saxena
            Patryk Diak
            Votes:
            0 Vote for this issue
            Watchers:
            6 Start watching this issue

              Created:
              Updated:
              Resolved: