This project automates the task of simulating and comparing different models. Classes were kept generic but have to be customized to be useful
- A minimal interface that should be implemented from all simulation classes is defined in simulation_base.py
- Factory method of simulation.py automates initialization of simulation classes which can have any number of custom named fields.
- Two helper classes automate the generation of time based plots and plots with respect to a custom parameter.
- Another class automates the storage of simulation data such that their plots can be generated at a later time without running the simulation again.
The functionality of simulator is demonstrated through the
implementation of two trivial cluster scheduling models
whose details are not in scope of this work.
They are in simulator.simulation_type1.py and
simulator.simulation_type2.py.
We will simply mention that both models serve two job types
and metric 1 corresponds to the number of jobs of type 1
in the system, while
metric 2 corresponds to the fraction of jobs of type 1
that were not rejected.
To test them we implemented two experiments in
the experiments folder. The experiment_simple.py
plots metric 1 with respect to time.
The experiment_series.py plots metric 2 at the end
of simulation with respect to model parameters like
the number of servers and how frequently jobs
arrive.
Invoke python from root directory to run any of the scripts
e.g.
> python experiments/experiment_simple.py
Parameter PROGRAM chooses whether to plot or run one of
the two examples. Parameter SAVE saves outputs after
running simulations (ignored if we choose to plot).
Parameter RUNTAG is supposed to be a unique identifier
of a run. It has to change in every run, if parameters are
different.
The design allows the addition of multiple simulation classes as long as they all have the same interface. This simplifies the comparison but not the implementation of individual classes. Also documenting each class and documenting experiments (e.g. which parameters or example was run for a specific runtag) are necessary and unavoidable.