Repository navigation
benchmarking: what to set and watch before a dense run - #2499
Conversation
0cc3c75 to
067bce6
Compare
|
Thank you for this! Have a couple of questions for clarity, although the general idea of a scaling guide is a good one. haiyanmeng you may find this interesting. |
haiyanmeng
left a comment
There was a problem hiding this comment.
Thanks for the contribution.
May I ask how did you identify these best practices? Did you encounter these bottlenecks while running the benchmark tests in a dense mode?
067bce6 to
30955cc
Compare
Yes, exactly what i did haiyanmeng some of what i did i am slowly pushing to https://github.com/dims/gym-substrate FYI |
Max Smythe (maxsmythe)
left a comment
There was a problem hiding this comment.
Added comments... forgot that they were in draft mode.
|
closing and reopening to re-trigger a flaky ci test |
d134a37 to
8b20fa5
Compare
808afb7 to
572a5d5
Compare
Max Smythe (maxsmythe)
left a comment
There was a problem hiding this comment.
LGTM, thank you for this!
Giving haiyanmeng a chance for any follow up comments before merging.
bee26e9 to
b78308d
Compare
Limits a dense run meets that the demo path never reaches: port-forward under load, PostgreSQL slots and probes, the parking budget, synchronized bursts, the neighbor table, restores under full load, the kubelet. Signed-off-by: Davanum Srinivas <[email protected]>
The pool flag has no default of 20: unset, pgx sizes the pool at the node CPU count. The port-forward point applies only to a load generator outside the cluster. Only the liveness probe restarts postgres. Signed-off-by: Davanum Srinivas <[email protected]>
Name the oversubscription behind "2x", say headroom rather than steering restores (Substrate has no such control), and spell out that the postgres probe fix includes the container resource requests. Signed-off-by: Davanum Srinivas <[email protected]>
b78308d to
72a84c8
Compare
|
Merging... doc only change. |
A dense run meets limits the demo path never reaches. One new section lists them with the setting for each: reaching the API and router without
kubectl port-forward, PostgreSQL slots and probe timeouts, the parking budget, synchronized bursts against the ext_proc breaker, the kernel neighbor table, restores under full load, kubelet CPU reservation, and what to watch during the run.Docs only.