-
Bug
-
Resolution: Done
-
Major
-
None
-
None
-
None
-
8
-
False
-
-
False
-
rhel-9
-
None
-
rhel-net-ovs-dpdk
-
-
-
ssg_networking
-
OVS/DPDK - FDP-25.E - 1, FDP-OVS/DPDK Sprint 7
-
2
In RHOSP 17.1 system, customer has a setup that is using 2 x F5 Load Balancer instances (vLB's) in Active/Standby mode:
host dpdk-0 <<< F5 vLB A
host dpdk-1 <<< F5 vLB B
Host dpdk-2 is used for iXIA Servers.
Host dpdk-3 is used for iXIA Clients.
All hosts are set up with option pmd-maxsleep.
iXIA Clients are sending traffic (and getting replies) to iXIA Servers via the active vLB.
The active vLB is owner of a "masquerade MAC".
Upon switchover, the newly active vLB becomes owner of the masquerade MAC and issues a GARP to announce this change.
Masquerade MAC is an F5 proprietary mechanism which is somewhat the MAC equivalent of a virtual IP address (VIP).
In switchover test, it takes 5 minutes before we can see traffic getting directed to the nely active vLB.
When pmd-maxsleep is removed, failover scenario works as expected
^^^^^^^
Note that this looks very much like https://bugzilla.redhat.com/show_bug.cgi?id=2213892
with the difference that pmd-maxsleep is enabled.
How reproducible:
Always
Steps to Reproduce:
Force vLB switchover.
Actual results:
What is observed on host dpdk-3 (hosting iXIA Clients):
- using ovs-tcpdump, we see the GARP from newly active vLB is received by OVS DPDK.
- OVS FDB table doesn't get updated.
What is observed on hosts dpdk-0 and dpdk-1:
- the vLB that becomes active issues the GARP
- FDB tables on both hosts get updated with new masquerade MAC port info.
Expected results:
What is expected is that upon receiving the GARP, OVS FDB tables on all compute hosts get updated.
Additional info:
RHOSP 17.1.4
openvswitch3.3-3.3.0-62.el9fdp.x86_64
other_config :
{dpdk-extra=" -n 4", dpdk-init="true", dpdk-socket-limit="4096,4096", dpdk-socket-mem="4096,4096", emc-insert-inv-prob="0", pmd-auto-lb="true", pmd-auto-lb-improvement-threshold="80", pmd-auto-lb-load-threshold="50", pmd-cpu-mask="2a8000000000002a8", pmd-rxq-assign=group, pmd-rxq-isolate="false", shared-mempool-config="9100:0,9100:1", smc-enable="true", vhost-postcopy-support="true"}Note that pmd-maxsleep is set/unset (on all 4 hosts) from the command line:
~~~
Mar 31 18:02:54 tpa-vim-c-computedpdkcl-3 ovs-vsctl[632387]: ovs|00001|vsctl|INFO|Called as ovs-vsctl set Open_vSwitch . other_config:pmd-maxsleep=220
Mar 31 18:47:20 tpa-vim-c-computedpdkcl-3 ovs-vsctl[914285]: ovs|00001|vsctl|INFO|Called as ovs-vsctl remove Open_vSwitch . other_config pmd-maxsleep
~~~
- is related to
-
FDP-1562 OVS learn action as configured by ML2 not sensitive enough to garp/nd packets during high throughput
-
- Resolved
-