This project is a PoC/example to show automation of large-scale Cloud Infra optimization with the orchestration tech and BO to decouple cross-domain expertise and accelerate the experiments to a new level.
Below is the arch digram:
Cloud-pipeline(README) automated expensive experiments by wiring the circle:
- With given parameter as input, Perform Parameter Applying towards the Cloud Infra System
- Running Backgroud Traffic load on the system with predefined Applications: code
- Instantiating a given Application under scheduling policy defined by the parameter and Run a Benchmark defined here
- Collect Benchmark data and cleanup env to be ready for next epoch
With the circle connected by Cloud-pipeline, we could run optimization with Bayesian Optimization in a jupyter notebook: demo
Those things put as predefined ones while should be inputs of the toolchain itself:
- The given parameter to tune
- The background traffic
- The optimization target for application workload & its Benchmark(desc here)
Also, a dashboard(code: frontend, backend) to help visualization of the training process and the outcome of the tuning was created below is a screen record for it: https://vimeo.com/497995660
And, not just calling them in python, a CLI for cloud-pipeline was also created to easily debug, operate the experiment in a handy way, here is a screen record for that: https://vimeo.com/497997340