| *** rosmaita1 is now known as rosmaita | 13:20 | |
| *** vhari_ is now known as vhari | 14:14 | |
| sean-k-mooney | do we have a better solution then rejecting a host in a pre playbook to ensure a job runs on a vm with 8vcpus | 16:35 |
|---|---|---|
| opendevreview | Merged openstack/project-config master: Stop publishing collection to PyPi https://review.opendev.org/c/openstack/project-config/+/978084 | 16:35 |
| sean-k-mooney | nova and cyborg have feature that reallly do need more then 8vcpu to be able to test | 16:35 |
| sean-k-mooney | i think the nested virt lables happen to be 8vpcu by aciddnet but that not actully the reqiurement | 16:36 |
| clarkb | sean-k-mooney: we are already rejecting nodes with only 1 vcpu | 16:37 |
| sean-k-mooney | right but the job can fail with the retry limit | 16:37 |
| sean-k-mooney | so that not really relibale | 16:37 |
| clarkb | that is the xen bug which can be worked around with a kernel boot parameters but it will impact every context switch the cpus do for all jobs on that image (ubuntu noble) | 16:38 |
| clarkb | then the new rax flex region instances only have 4 vcpus but that has been deemed fine as they are zoomy | 16:38 |
| sean-k-mooney | right but that not actully fine for some nova feature | 16:38 |
| sean-k-mooney | like cpu pinning | 16:38 |
| sean-k-mooney | or some feature we want to test in cybrog | 16:38 |
| fungi | the nested virt label avoids rackspace legacy, and therefore the xen bug, but yeah doesn't avoid 4cpu nodes in flex | 16:38 |
| clarkb | right the only way to properly fix that is to either accept a performance decrease across the board on that image with nox2apic. Stop using the cloud region entirely (its like half our quota). Or stop using newer kernels | 16:39 |
| sean-k-mooney | that will use kernel moduel where we we need ot allcoate a core to the kernel module | 16:39 |
| sean-k-mooney | clarkb: so the usecause im talkign about is unrelated to the xen issue | 16:39 |
| clarkb | we;ve never promised 8 vcpus https://docs.opendev.org/opendev/infra-manual/latest/testing.html we call out different cpu copunts and an explicit difference between nodes | 16:40 |
| sean-k-mooney | im well aware | 16:40 |
| fungi | if this is something that we want to solve, then a custom label limited to flavors with 8 cpus or more would be the most efficient solution | 16:40 |
| sean-k-mooney | im stating that i would like a way to request >6 vcpu or spcificly 8 | 16:40 |
| clarkb | right if the issue isn't the xen issue but rax flex vs everything else then we'd need different labels | 16:41 |
| fungi | running jobs on the 16gb or 32gb labels might work around it, but will consume more memory quota obviously | 16:41 |
| clarkb | that only map to clouds where there are 8vcpus. There isn't a way to use 4vcpu? allocating one cpu still leaves you 3 | 16:41 |
| sean-k-mooney | right so im wondierign if it would be ok to createa single new lable ubuntu-noble-8vgpu-8G | 16:41 |
| sean-k-mooney | and use that just in the job or jobs that woudl need it | 16:42 |
| sean-k-mooney | clarkb: we need 2 core for the nvme emulator and we need to use isocpus | 16:42 |
| sean-k-mooney | meain runing devstack with effeictly 2 core for eveything else | 16:43 |
| sean-k-mooney | im currently wondering if we coudl make this a 2 node job with a deicated compute | 16:43 |
| sean-k-mooney | that might work | 16:43 |
| sean-k-mooney | for https://review.opendev.org/c/openstack/cyborg/+/977083/9#message-13a90c9ce5b990729858bc6a80f53a254589a841 | 16:43 |
| clarkb | yes 2 node should double the total cpu count if you can slice things effectively across the nodes | 16:44 |
| sean-k-mooney | for feature like cpu pinnign haveing 4 cores total is hard | 16:44 |
| sean-k-mooney | well a compute only has nova/neutorn libvirt and ovs/ovn | 16:45 |
| sean-k-mooney | so that proably woudl be workable | 16:45 |
| sean-k-mooney | we coudl perhaps test cpu pinning in nova with 4 cores if concrrancy was 1 | 16:45 |
| sean-k-mooney | so that coudl be an option as well but we are really jumping through hoops to make it workable | 16:46 |
| sean-k-mooney | fewer faster cores is proably fine for most services but it ligitmatly inviladge or limits testing for some projects | 16:47 |
| sean-k-mooney | its why we historically have not tested these feature in the first party ci in the past | 16:47 |
| opendevreview | Michal Nasiadka proposed openstack/project-config master: propose-updates: Add test-requirements target https://review.opendev.org/c/openstack/project-config/+/978566 | 16:52 |
| sean-k-mooney | the nvmevirt repos say "t is highly recommended to use the isolcpus Linux command-line configuration to avoid schedulers putting tasks on the CPUs that NVMeVirt uses:" | 16:54 |
| sean-k-mooney | but i wonder if we coudl get away with not doing that | 16:54 |
| sean-k-mooney | as long as its fucntional enough for ci im not sure we care about the performance | 16:55 |
| sean-k-mooney | sicne most test wont use the virtualised nvme device | 16:55 |
| sean-k-mooney | only the one that is testing that feature | 16:55 |
| sean-k-mooney | we migt get away with oversubcibign those cores | 16:55 |
| *** mtreinish_ is now known as mtreinish | 17:30 | |
| *** mtreinish_ is now known as mtreinish | 17:53 | |
Generated by irclog2html.py 4.1.0 by Marius Gedminas - find it at https://mg.pov.lt/irclog2html/!