search

Rhodes-CS-Department / jupyterhub-deployment

Configs and READMEs for Rhodes CS JupyterHub deployment
2 stars 1 forks source link

Get a better characterization of CPU/RAM use per notebook #28

Open ml8 opened 2 years ago

ml8 commented 2 years ago

It looks like we are still fairly overprovisioned:

Screen Shot 2021-10-26 at 11 54 18 AM

Since CPU use looks bursty, we could probably reduce the CPU reservation dramatically and let the cap stay high. Memory utilization looks like ~200MB in the steady state. Perhaps we could downsize to 500MB reservation and 1G cap. We want to be slightly overprovisioned in memory, since exceeding the CPU cap just makes a notebook slow, vs. exceeding the memory cap will lock all notebooks on the server until the user manually kills notebooks.

@trenthem is this a concern to you at all, or would further reduction in VM time (e.g., 20%) just be noise/not worth the person-hours to tune?

We reduced usage by about 50% before we launched this semester, for context.

ml8 commented 2 years ago

Over Fall'21 each server was still overprovisioned. Leaving this open if @anesepark or @walidabualafia want to use stackdriver metrics to get a better characterization of exactly how overprovisioned we are.