Run vBNG VNF/CNF tests and collect results

[lixuna]

Tasks:

• Close down all VM and prepare system for running test
• Run VNF test and collect results
• Run CNF test and collect results

[michaelspedersen]

vBNG VNF Benchmarks

• Test configuration:

  • L2 Frame: 68B
  • Duration: 60 seconds
  • Load: 3Mpps (limited by Pktgen)
  • Iterations: 5
  • CPU pinning: 1-3 (Host VPP), 4-6 (Pktgen), 7-9 (vBNG)

• Results:

  • Throughput
    • 2,960,379 pps
    • 2,939,770 pps
    • 2,984,178 pps
    • 2,940,379 pps
    • 2,977,720 pps
    • Avg: 2,960,485 pps
  • Memory
    • Resident Set Size (RSS): 77.91 MB (76088 KiB)

CPU pinning done on running VMs using virsh list followed by virsh vcpupin <id> <vcpu> <cpu> for both Pktgen and vBNG. Host VPP is pinned and isolated through kernel parameters and VPP configuation.

Performance is currently limited by Pktgen, as each of the two ports only support 1 rx/tx queue, which limits the CPU usage to 2 cores. This limitation will be true for both VNF and CNF tests as both will use the same VM version of Pktgen. Once we figure out an alternative to the vhost-user interfaces (e.g. SR-IOV) it should be possible to increase the queues per port, which should increase the throughput and remove the bottleneck.

[michaelspedersen]

vBNG CNF Benchmarks

• Test configuration:

  • L2 Frame: 68B
  • Duration: 60 seconds
  • Load: 3Mpps (limited by Pktgen)
  • Iterations: 5
  • CPU pinning: 1-3 (Host VPP), 4-6 (Pktgen), 7-9 (vBNG)

• Results:

  • Throughput

    • 2,931,671 pps
    • 2,951,497 pps
    • 2,834,269 pps
    • 2,921,445 pps
    • 2,954,347 pps
    • Avg: 2,918,646 pps

  • Memory

    • Resident Set Size (RSS): 22.4 MB

The difference in throughput between VNF and CNF is negligible, as both are limited by the number of packets generated by the Pktgen VM

Performance is currently limited by Pktgen, as each of the two ports only support 1 rx/tx queue, which limits the CPU usage to 2 cores. This limitation will be true for both VNF and CNF tests as both will use the same VM version of Pktgen. Once we figure out an alternative to the vhost-user interfaces (e.g. SR-IOV) it should be possible to increase the queues per port, which should increase the throughput and remove the bottleneck.

[lixuna]

This ticket can be closed