Initialisation script to set up VM from cold start
Just having a nicely configured VM doesn’t anwer the whole question in a machine learning scenario. In particular,
using preemptible machines means that the VM is hard-stopped at minimum every 24hrs, and a
typical way to do training, etc, involves setting up
screen to allow for clear disconnects.
So, in addition to mounting an extra drive, wouldn’t it be nice to get some screen sessions going with the correct paths,
virtualenv already running, and ready-to-roll?
Create a startup script locally
Create this script as a local file
startup.bash (clearly, your details will be very different,
but some essential elements are included here) :
Create a ‘startup hook’ for the GCP machine
Post that script to the VM using the
gcloud add-metadata command from your local machine :
Test that it works
Go on! Just
start an instance and see whether it works!
If you need to change the script, just make the changes locally, and redo the
add-metadata step : It
seems to replace what was there before. Then, you’ll have to
start the instance again -
best done during down-time, rather than when hot on the trail of the latest model…
blog comments powered by Disqus