From f771af4e69db4b8937f64fbf4024eb518a7cc230 Mon Sep 17 00:00:00 2001 From: mjkwiatkowski Date: Mon, 1 Jun 2026 14:15:37 +0200 Subject: feat: added some notes --- notes/auto-scaling.txt | 43 +++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 43 insertions(+) create mode 100644 notes/auto-scaling.txt (limited to 'notes/auto-scaling.txt') diff --git a/notes/auto-scaling.txt b/notes/auto-scaling.txt new file mode 100644 index 0000000..a6b009f --- /dev/null +++ b/notes/auto-scaling.txt @@ -0,0 +1,43 @@ +==AutoScaling== +To get it to run --? look at how failures are implemented. +The moment your pool of usable hosts becomes too tight then you activate new ones. +1) I want to implement which allows me to put hosts inactive and activate it again. +2) During the workload the operator should be able to close and open the workloads. +i) then you build a policy as to when you activate new hosts and when you want to stop hosts. +ii) have many policies to do this +iii) add booting time --> cold start. A host becomes available but it takes still more time to make it workable because you need to boot everything. +iv) cold starts use zero power or very minimal --> idle power of a node that is turned off. Inactive have minimal power draw, but the downsides is that nodes may not be available immediately. +Inactive takes more time to boot e.g., 5 minutes. +The more you can predict how many nodes you will need to more you can save power without impacting the performance. + +The bachelor focus -- the primary focus should be on the design of the digital twin. +Be careful that you do not make implementing the main topic of the bachelor thesis. +The main topic should still be digital twinning. + +The digital twin is activated in 2 ways: either the operators prompt the digital twin when they run a workload or the digital twin sends a notification to the datacenter. + +IBM Dublin := you have metrics being managed and policies can decide to do actions. +And these actions can be many different things. +Currently does not matter what it would -- this could be auto-scaling, scheduling, routing. +What is the difference between routing and scheduling? + +Great success == we do not include AI inference in the bachelor thesis! +Policy decisions can be made using different heuristics --> AI inference. +Do not work too much on the policies. + +The focus of the thesis is the digital twinning part. + +Autoscaling vs. failures vs. scheduling. +MVP soon. + +Why can't we port a digital twin from other domains? +Show that a digital twin can react. + +Show the readers what would the perfect experiment look like in the perfect. + +What a digital twin is --> give the answer in the background somehow. +What do you think a digital twin is? Make it clear. +Why is a specific datacenter digital twin different from what there already is. + +Versen Thesis Awards they are promoted at ICTO today and tomorrow. + -- cgit v1.2.3