Slow starting time compared with lite

[bisgaard-itis]

This is a placeholder for https://z43.fogbugz.com/f/cases/193514/Slow-startup-times-of-sim4life-io

Tasks

Preview Give feedback

1. Autoscaling: use labels instead of draining nodes osparc-simcore#5339
   a:autoscaling a:director-v2 +2
   
2. ✨Autoscaling: use label instead of draining machine osparc-simcore#5340
   a:autoscaling +1
   
3. 🐛 Autoscaling: issues with labelled drained nodes osparc-simcore#5348
   a:autoscaling +1
   

[YuryHrytsuk]

Osparc ops issue:

• Slow starting time compared with s4l-lite osparc-ops-environments#552

The problem is explained in details: ITISFoundation/osparc-ops-environments#552 (comment)

[sanderegg]

facts:

• s4l-lite uses a shared model (i.e. 3/6 s4l-lite run on a single machine)
• s4l uses an exclusive model (i.e. 1 s4l on a single machine)
• cold/fresh machine: is started on demand, its boot disk is EBS-based and initial access is very slow as all the data is lazily pulled from AWS-backend, GPU machines typically provide a instantly fast SSD drive as well, but it is always empty.
• images pre-pulling: specific docker images are pulled from a registry at machine initial start
• hot/active machine: is a machine where all the necessary images were already pulled once and has running s4l-lite or s4l services
• warm/buffer machine: is a started machine where image pre-pulling was achieved to improve initial starting time of s4l-lite or s4l services

[YuryHrytsuk]

It was fixed by 2 PRs below. The fix is working so far. We will observe further before closing.

• ✨Autoscaling: use label instead of draining machine osparc-simcore#5340
• 🐛 Autoscaling: issues with labelled drained nodes osparc-simcore#5348

[sanderegg]

case is completed