| *** tobberydberg_ is now known as tobberydberg | 11:51 | |
| mgariepy | jamesdenton, https://bugs.launchpad.net/nova/+bug/1751923 | 12:39 |
|---|---|---|
| jamesdenton | nice find! i was wondering in that instance_info_cache was related | 12:50 |
| jamesdenton | we have run into this before | 12:50 |
| jamesdenton | mgariepy i found that my horizon issue last week was related to having two (different) cells named cell1. Not ssure how or when that creeped in | 12:55 |
| mgariepy | yep. the fun part is that it's not merged for Rocky lol | 12:56 |
| spatel | jamesdenton morning | 13:10 |
| jamesdenton | hi | 13:12 |
| mgariepy | jamesdenton, after some digging.. i only had 3 vms with the issue. | 13:23 |
| mgariepy | one which i fixed manually (not realizing it had the issue. | 13:24 |
| spatel | jamesdenton do you know how to do vm bonding when using two sriov vf ? | 13:26 |
| jamesdenton | i recall seeing a blueprint for that functionality but i don't know if it has been implemented. It would require the ability to schedule two VFs across two PFs, which i don't think exists (yet)? | 13:28 |
| jamesdenton | you should be able to give a VM two interfaces from the same network and bond them in the VM? | 13:28 |
| spatel | Yes that is what i am thinking | 13:36 |
| spatel | jamesdenton but i would like to have some good way and not hack way... | 13:37 |
| jamesdenton | i don't know that there is a "sanctioned" way to do it | 13:37 |
| spatel | when i say give create vm then it will automatically pic two VF from two PF and attach to vm and then cloud-init will take care of bonding | 13:37 |
| jamesdenton | last i knew, the behavior was "pick two VF, likely from same PF until that PF is exhausted" | 13:38 |
| spatel | This guy did it but process is not joyful - https://www.redpill-linpro.com/techblog/2021/01/30/bonding-sriov-nics-with-openstack.html | 13:38 |
| jamesdenton | yeah, ok, two different networks but really the same network. that;'s also an option :D | 13:39 |
| spatel | yes.. | 13:40 |
| spatel | but its hard to put them in terrafrom :( | 13:40 |
| spatel | we need way to automate it so my clients can build vm without lots of hacks | 13:41 |
| jamesdenton | understood. were you ever able to make headway on the DPDK tuning? | 13:41 |
| spatel | still running load-test | 13:42 |
| spatel | comparing sriov with dpdk | 13:42 |
| spatel | majority of people running dpdk load-test with testpmd inside guest VM | 13:42 |
| spatel | testpmd is dpdk aware application but in real world we don't do testpmd, we run raw application which is not dpdk aware | 13:43 |
| spatel | for example haproxy | 13:43 |
| spatel | feel like i am running wrong way load-test and that is why my result is different compare to people running with testpmd | 13:44 |
| jamesdenton | i didn't think the VM needed to be DPDK aware if it was happening on the compute, but dunno | 13:46 |
| spatel | do you have dpdk environment ? i would like to see your result also | 13:46 |
| spatel | let me show you result | 13:48 |
| jamesdenton | i do not | 13:48 |
| spatel | jamesdenton i have two vm 1. sriov and 2. dpdk and here is the result | 13:49 |
| spatel | https://paste.opendev.org/show/809694/ | 13:49 |
| spatel | This is UDP based iperf3 load-test | 13:49 |
| spatel | with packet size 150bytes | 13:49 |
| spatel | jamesdenton one thing i have observed which on sriov vm context switching is less and on dpdk vm its 3x higher | 14:02 |
| spatel | may be because its virtio and packet routing via kernel but in sriov its direct VF so packet traveling without using kernel | 14:03 |
| jamesdenton | yeah | 14:04 |
| jamesdenton | how many PMDs and how much RAM did you allocate? | 14:04 |
| jrosser | noonedeadpunk: i am wondering if we have a mistake, this https://github.com/openstack/openstack-ansible/blob/master/playbooks/haproxy-install.yml#L46-L51 and https://github.com/openstack/openstack-ansible-haproxy_server/blob/master/tasks/main.yml#L34-L49 at the same time? | 14:07 |
| spatel | 1G and 4 pmd | 14:08 |
| noonedeadpunk | jrosser: hm... I think https://github.com/openstack/openstack-ansible/blob/master/playbooks/haproxy-install.yml#L46-L51 does smth different? | 14:09 |
| noonedeadpunk | but not sure... | 14:09 |
| noonedeadpunk | because we provide only pki_search_certificates_pattern | 14:09 |
| noonedeadpunk | So we shouldn't generate certs there | 14:09 |
| noonedeadpunk | But I'd say that https://github.com/openstack/openstack-ansible-haproxy_server/blob/master/tasks/main.yml#L34-L49 should cover usecase anyway | 14:10 |
| noonedeadpunk | where we have issue is https://bugs.launchpad.net/openstack-ansible/+bug/1938961 | 14:10 |
| noonedeadpunk | I started testing out solution but had to switch :( | 14:10 |
| opendevreview | Dmitriy Rabotyagov proposed openstack/openstack-ansible-os_keystone master: Fix apache ProxyPass arguments https://review.opendev.org/c/openstack/openstack-ansible-os_keystone/+/811402 | 14:11 |
| opendevreview | Jonathan Rosser proposed openstack/openstack-ansible-haproxy_server master: Fix typo for user supplied certificate variable https://review.opendev.org/c/openstack/openstack-ansible-haproxy_server/+/811979 | 14:17 |
| opendevreview | Dmitriy Rabotyagov proposed openstack/openstack-ansible-haproxy_server master: Fix haproxy Let's Encrypt SSL path https://review.opendev.org/c/openstack/openstack-ansible-haproxy_server/+/811985 | 14:48 |
| *** promethe- is now known as prometheanfire | 15:11 | |
| *** prometheanfire is now known as Guest1415 | 15:12 | |
| *** chandankumar is now known as raukadah | 15:21 | |
| *** Guest1415 is now known as prometheanfire | 15:35 | |
Generated by irclog2html.py 2.17.2 by Marius Gedminas - find it at https://mg.pov.lt/irclog2html/!