serverless-measurement

Measurement code for Serverless-based container life-cycle

Purpose:

Built specifically to deploy and measure Knative serverless function's lifecycle in terms of energy, resource consumption and latency. However, it can be used for any K8S-related platform to conduct the same purpose. This code is able to collect:

Resource consumption -- CPU, GPU, RAM (%)
Power consumption (W) 
Energy consumption (Wh/J)
Duration (s) and response time (ms)
Frame Per Second (FPS) for app that exposes API to get this info

Before running:

You can see the required library in the main.py files, other from that, there are some prerequisites as follows: ⋅⋅* File curl.yaml must be deployed and stay in the system forever, it serves as the container for curl cmd to run over ⋅⋅* Prometheus must be installed and its address must be given in the variable file ⋅⋅* Tinkerforge devices and its relevant libs must be implemented in order to be used the power.py file

Basic cmds:

File auto_generate.py contains a loop, which repetitively runs the main.py file. Why repetition? because it's basic measurement requirement. Run auto_generate.py by: python3 auto_generate.py

File main.py control which part of lifecycle needs to be measured, it can be run independently for testing purpose. This files has three important functions:

1. collect_life_cycle for measuring most of the lifecycle.
2. curl_latency for measuring the response time perceived by end-users through curl cmd
3. collect_cold_warm_disk for temporarily measuring the Cold to Warm Disk process. This is previously measured by detecting the Pulled status of deployment

However, a bug occurred in the code leading to wrong detection. Notice that result from this process must be minused with Null to Cold and Warm Disk to Warm CPU to get the real Cold to Warm Disk value.

Since Knative is having problem with terminating deployment, an auto_delete process is written inside functional_methods to automatically detect Terminating pod's status and send a terminated signal there. This process automatically runs when user runs file auto_generate.py.

In order to manually delete one or multiple pods that are stuck at Terminating without running measurement, please run python3 delete.py. This function acts just like the described one above, but as a standalone version.

Advanced cmds: this has been removed!

There is a separated process called freeze mem, which measures the freeze memory state of the serverless function. To measure this process, the following order must be complied: 1. Remove every deployment that are currently running 2. Apply files and change config following this link 3. Use the main_mem_freeze.py file instead of the main.py* 4. After finishing freeze mem measurement, the steps 1 and 2 above must be reversed. > * *currently freeze mem can only measure its state, its actions are measured by curl tool *