Task: Right size a VM by profiling a job¶

Use-cases:

Reduce costs by using smaller VMs that fit the workload better
Increase density by running more jobs on the same hardware
Improve performance bottlenecks by adding more resources when needed

In the blog post Right sizing VMs for GitHub Actions, we introduce the need for profiling (or metering) a running job in order to make sure the VM is the right size for the workload.

A range of metrics are collected in addition to the standard ones like CPU & RAM consumption, vmmeter also shows contention on I/O, whether a job is running out of disk apce, and how many open files are in use. Non-obvious metrics like entropy and I/O contention are also collected, which can also be linked to degraded performance.

Try it out¶

Add the following to the top of your worklflow:

steps:
# vmmeter start
        - uses: alexellis/setup-arkade@master
        - uses: self-actuated/vmmeter-action@master
# vmmeter end

The vmmeter-action will run in the background and collect metrics about the job. Its Post run action will collect the metrics and upload them to the job's summary.

Here is an example from building a Linux Kernel:

Total RAM: 61.2GB
Total vCPU: 32
Load averages:
Max 1 min: 5.76 (18.00%)
Max 5 min: 1.34 (4.19%)
Max 15 min: 0.44 (1.38%)


RAM usage (10 samples):
Max RAM usage: 2.348GB


Max 10s avg RAM usage: 1.788GB
Max 1 min avg RAM usage: 1.233GB
Max 5 min avg RAM usage: 1.233GB


Disk read: 405.4MB
Disk write: 469.6MB
Max disk I/O inflight: 0
Free: 45.16GB   Used: 4.66GB    (Total: 52.52GB)


Egress adapter RX: 275.6MB
Egress adapter TX: 1.341MB


Entropy min: 256
Entropy max: 256


Max open connections: 23
Max open files: 1856
Processes since boot: 19819


Run time: 47s