MAAS deployments on Grace platforms should not be so slow
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
maas-images |
In Progress
|
High
|
dann frazier |
Bug Description
MAAS deployments on NVIDIA Grace-based platforms currently take a long time. On a Supermicro Grace/Grace board, the deployment time was 46 minutes - and often they will time out. I found that the deployment process is often CPU-bound. And this slowness only seems to be an issue during MAAS deployments. After some investigation, the significant difference I found to be the presence of the cppc-cpufreq kernel module. If I download and manually load that module early in the deployment phase, things speed up noticeably - the last such deployment took just 20 minutes. I suspect the deployment time would be cut down even more if the module were loaded earlier than I'm able to do manually - as that can only happen after networking/ssh keys are ready.
My hypothesis is that the cores on this SOC default to a low-frequency unless this module is available. I can not demonstrate this because the frequency attributes in sysfs are only available when this module is loaded.
Related branches
- Jack Lloyd-Walters: Approve
-
Diff: 12 lines (+1/-0)1 file modifiedbin/kpack-from-image (+1/-0)
summary: |
- MAAS deployments on Grace platforms are very slow + MAAS deployments on Grace platforms should not be so slow |
Changed in maas-images: | |
status: | New → In Progress |
importance: | Undecided → High |
assignee: | nobody → dann frazier (dannf) |