Auto-scaling is the capability of Function as a Service systems, that supports dynamic
scaling of the function instances according to the incoming load. Auto-scalers fire
scaling events when a certain threshold is exceeded. However, if this threshold is
not set properly, the function can suffer from under or over-provisioning. In this
paper we introduce an autoscaling solution for compute intensive functions that calculates
the scaling threshold according to the user needs and keeps the completion times predictable
even when the function is scaled out. The scaling threshold is given by a simulator
that determines the completion time distribution of the function for a given load.
We also show which load-balancing algorithm is recommended to use for our auto-scaler.
We compare our auto-scaler to existing ones implemented in open source serverless
projects.
