Autoscaling according to memory vs CPU

[Tianyi Wu]

Dear Google cluster data researchers,

I saw in section 3.2 of the autopilot paper that jobs are being scaled according to memory or CPU:

"The Autopilot service chooses the recommender(s) to use for a job according to whether the resource being autoscaled is memory or CPU; how tolerant the job is to out-of-resource events (latency tolerant or not, OOM-sensitive vs. OOM tolerant); and optional user inputs, which can include an explicit recommender choice, or additional parameters to control Autopilot’s behavior, such as upper and lower bounds on the limits that can be set.  "

I was wondering if this means tthat the trace datas are generated based on memory since the papers focused on memory and OOM.

Best,

Tianyi