#opnfv-ovsnfv: Weekly Open vSwitch for NFV meeting (28th November 2016)

Meeting started by mdgray_ at 14:03:02 UTC (full logs).

Meeting summary

    1. Mark Gray (mdgray_, 14:03:15)

  2. Roll Call (mdgray_, 14:03:24)
    1. Mark Gray (mdgray_, 14:03:27)
    2. Chunghan Lee (lch, 14:03:28)
    3. Tom Herbert (tfherbert, 14:03:41)
    4. Chunghan Lee (lch, 14:03:55)
    5. Billy O'Mahony (billyo_, 14:03:56)

  3. Readout on the priority path development work (Billy) (mdgray_, 14:04:54)
    1. Priority path use case is that we need to enable some traffic class with better latency characteristics that others. (mdgray_, 14:06:52)
    2. Billy presents tests done to measure the current behaviour in OVS. (mdgray_, 14:07:16)
    3. Billy used three traffic classes (Bulk data transfer, Voice, PTP) (mdgray_, 14:08:35)
    4. Billy presented latency data for different traffic classes (mdgray_, 14:11:38)
    5. As the PTP rate is so low it causes a sampling error. There are not enough PTP frames to show the long tail in latency. This is currently speculation. Billy will confirm this. (mdgray_, 14:15:10)
    6. These tests were a PVP test. (mdgray_, 14:15:45)
    7. Billy outlined what the graphs should look like. i.e. Latency spikes are only seen in the lower priority traffic (mdgray_, 14:23:12)
    8. Billy also opened a discussion on the ovs-discuss mailing list (mdgray_, 14:23:27)
    9. https://mail.openvswitch.org/pipermail/ovs-discuss/2016-November/043081.html (mdgray_, 14:23:38)
    10. Billy will share excel spread sheet with results (mdgray_, 14:24:12)

  4. Colorado Fuel Performance Results (Billy) (mdgray_, 14:24:21)
    1. Billy has seen a performance result in yardstick tests between Brahmaputra and Colorado (mdgray_, 14:24:49)
    2. Yardstick pktgen test shows that although ovs-dpdk performs better than vanilla ovs across Brahmaputra and Colorado. However, ovs-dpdk performance less better in colorado than brahmaputra. (mdgray_, 14:30:22)
    3. Ping latency test looks about the same for colorado and brahmaputra. (mdgray_, 14:30:52)
    4. netperf tests (which are not part of yardstick) also showed decrease performance. (mdgray_, 14:32:17)
    5. Billy has not identified the root cause. (mdgray_, 14:33:00)
    6. Current assumption is the Qemu version is causing problems with virtio. (mdgray_, 14:33:24)
    7. Brahmaputra was using a different version of Qemu. (mdgray_, 14:34:16)
    8. Currently, we do not have results for Apex in order to compare. (mdgray_, 14:35:42)
    9. This is because the Apex infrastructure in that POD is using Cisco UCS which is adding a vlan tag which is not needed. Tom has worked around this by adding a rule. However, we need a more permanent solution. Some solutions were discussed. (mdgray_, 14:52:56)
    10. There is a similar issue in the FDS project. (mdgray_, 14:53:11)
    11. Some consensus to fix this in OVS. (mdgray_, 14:53:42)
    12. When Tom submits the patch, he will cc Billy in order to help move this through the OVS patch process. (mdgray_, 14:54:25)

  5. Other? (mdgray_, 14:55:21)


Meeting ended at 14:55:34 UTC (full logs).

Action items

  1. (none)


People present (lines said)

  1. mdgray_ (33)
  2. collabot` (4)
  3. lch (2)
  4. tfherbert (1)
  5. billyo_ (1)


Generated by MeetBot 0.1.4.