On Mon, Apr 25, 2022 at 8:23 PM 刘梦馨 <[email protected]> wrote: > > > Hi, Han > > I just found flow-restore-wait option in ovs-vswitchd.conf.db(5) can we just use it to simulate ovn-ofctrl-wait-before-clear to reduce upgrade downtime in old version ovn
It is possible, but depending on your use cases. There are some major differences between the two approaches: 1. With ovn-ofctrl-wait-before-clear, we don't need to restart OVS when upgrading ovn-controller, while flow-restore-wait works only with OVS restart. This means the operations for upgrading is much more complicated using flow-restore-wait. With ovn-ofctrl-wait-before-clear there is no special steps required during ovn-controller upgrade. 2. flow-restore-wait is more about keeping the old datapath flows untouched during the upgrade, but the ovs-vswitchd user space flows won't work, so this only works for packets that doesn't need any slow-path handling. For OVN this is not always true, e.g. ARP responders, ICMP handling, etc. would require ovs-vswitchd to participate. If those are unimportant during the upgrade (possibly very long due to large scale), then it is fine. 3. flow-restore-wait is particularly useful when packet miss (and upcall to slow path) is not acceptable during OVS/ovn-controller upgrade (probably for traffic that can't even tolerate the short latency increase during the rare situation of an OVS/OVN upgrade). ovn-ofctrl-wait-before-clear doesn't cover such use cases (but not conflict with flow-restore-wait, of course). Thanks, Han > > On Tue, 19 Apr 2022 at 03:23, Han Zhou <[email protected]> wrote: >> >> Whenever OpenFlow connection between ovn-controller and OVS is >> connected/reconnected, typically during ovn-controller/OVS >> restart/upgrade, ovn-controller would: >> 1. clears all the existing flows after the initial hand-shaking >> 2. compute the new flows >> 3. install the new flows to OVS. >> >> In large scale environments when there are a big number of flows >> needs to be computed by ovn-controller, the step 2 and 3 above can take >> very long time (from seconds to minutes, depending on the scale and >> topology), which would cause a data plane down time. >> >> This series solves the problem: patches 1 and 2 avoids the down time introduced >> by step 2; patches 3 and 4 avoids the down time introduced by step 3. >> >> A test shows zero ping loss (with 0.1s interval), while without the patches it >> was 100% ping drops during ~13s. >> >> Han Zhou (4): >> ofctrl: Wakeup when entering S_UPDATE_FLOWS. >> ofctrl: Support ovn-ofctrl-wait-before-clear to reduce down time >> during upgrade. >> ofctrl.c: Include group changes to bundle. >> ofctrl.c: Use bundle to avoid data plane downtime during the first >> flow installation. >> >> controller/ofctrl.c | 275 +++++++++++++++++++++----------- >> controller/ofctrl.h | 4 +- >> controller/ovn-controller.8.xml | 34 ++++ >> controller/ovn-controller.c | 6 +- >> tests/ovn-controller.at | 52 ++++++ >> 5 files changed, 275 insertions(+), 96 deletions(-) >> >> -- >> 2.30.2 >> >> _______________________________________________ >> dev mailing list >> [email protected] >> https://mail.openvswitch.org/mailman/listinfo/ovs-dev > > > > -- > 刘梦馨 > Blog： http://oilbeater.com > Weibo: @oilbeater _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
