On 11/19/24 03:23, Jun Wang wrote: > >> On 11/15/24 08:15, Jun Wang via discuss wrote: >>> We are using OVS bond with the configuration as shown below. We found that >>> when the switch is configured in MLAG mode: >>> >>> 1. We tested sending 1000 UDP packets per second, with a total of 500,000 >>>packets. The statistics show that there is occasionally a chance of losing a >>>few packets (background traffic is around 100,000 pps). >>> 2. We also tested the same scenario in a non-MLAG setup with the same >>>testing steps, and there was basically no packet loss (background traffic is >>>around 100,000 pps). >>> >>> >>> |ovs-vsctl --may-exist add-bond br-tun dpdk_tun_port tun_port_p0 >>> tun_port_p1 \| >>> |bond_mode=balance-tcp lacp=active >>> other-config:bond-rebalance-interval=||1000| >>> |other_config:lb-output-action=||true| |\| >>> |-- set Interface tun_port_p0 type=dpdk \| >>> |options:dpdk-devargs=||0000||:ca:||00.0| |mtu_request=||1600| |\| >>> | ||options:n_rxq_desc=||2048| >>> |options:n_txq_desc=||2048| |\| >>> | ||options:n_rxq=||4| |\| >>> |-- set Interface tun_port_p1 type=dpdk \| >>> |options:dpdk-devargs=||0000||:ca:||00.1| |mtu_request=||1600| |\| >>> | ||options:n_rxq_desc=||2048| >>> |options:n_txq_desc=||2048| |\| >>> | ||options:n_rxq=||4| >>> >>> For the phenomenon described above, in the MLAG scenario, after adjusting >>> |bond-rebalance-interval=0|, we found that during >>> the test of sending 1000 UDP packets per second with a total of 500,000 >>> packets, there was no packet loss. It is unclear whether >>> this is an issue related to OVS bond's support for MLAG, which might cause >>> packet loss under certain conditions. Therefore, >>> I found some relevant descriptions online that may explain this behavior. >>> >>> https://forum.proxmox.com/threads/solution-multiple-ovs-sdn-and-lacp-heartaches.120634/ >>> >>> <https://forum.proxmox.com/threads/solution-multiple-ovs-sdn-and-lacp-heartaches.120634/> >>> The description is as follows: >>> “Since my original post, I've spent some more time on this, and >>> discovered that OVS is totally unsuitable for more enterprise-y setups, >>> and MLAG in particular. None of the MLAG implementations I've tried this >>> with expect distinct flows to flap between legs, and this does cause minor >>> issues.“ >>> >>> In addition, the |bond-rebalance-interval=0| configuration can lead to >>> traffic imbalance between interfaces, which has a negative impact. Is there >>> a better solution to this issue? >> > >>If the switches do not support traffic moving between legs of the MLAG, >>then disabling of the rebalancing sounds like the only option. > > I have confirmed that the switch supports MLAG, but I occasionally encounter > packet loss, which is very frustrating and unacceptable. > >>Note that the traffic will still be balanced based on the packet hash. >>So, as long as you have many light traffic streams the traffic should >>be well balanced. But if your traffic pattern involves having a couple >>heavy streams along many light ones, then it may not be fully balanced >>indeed. > > I captured packets and confirmed that OVS is dropping packets. Observing the > coverage statistics, > I noticed an increase in |datapath_drop_invalid_port|. However, after setting > |bond-rebalance-interval=0|, the counter no longer increases.
Hmm, this is strange. Do you see ports go down intermittently? Is there something interesting in the ovs-vswitchd.log ? In general, all the operations with the bond table in the datapath are RCU-protected, so rebalancing should not cause any, even short, time frames where the port id is not valid. So, the increase of the counter is very strange. Also, what is your OVS version? > >>Is there a linux bond mode that works fine with this MLAG? >> >>> >>> ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ >>> Jun Wang >>> >>> _______________________________________________ >>> discuss mailing list >>> disc...@openvswitch.org >>> https://mail.openvswitch.org/mailman/listinfo/ovs-discuss > > ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ > Jun Wang > _______________________________________________ discuss mailing list disc...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-discuss
