converts Decima's actions to the environment's format

transforms environment observations into a format that's more suitable for Decima

Maintains all executor assignments. These include: - current location of each executor, called a 'pool' - commitments from one executor pool t

An object representing a job in the system, containing a set of stages with dependencies stored in a dag.

allow duplication in set

context manager which profiles a block of code, then prints out the function calls sorted by cumulative execution time

renders frames that visualize the job scheduling simulation in real time. A gantt chart is displayed, with the traces of all the workers are s

model updates occur at regular intervals, regardless of when the environment resets

model updates are synchronized with environment resets

Samples each episode's time limit from an exponential distribution

Original Decima architecture, which uses asynchronous message passing as in DAGNN. Paper: https://dl.acm.org/doi/abs/10.1145/3341302.3342080

Proximal Policy Optimization

Interface for all schedulers

A Gymnasium environment that simulates DAG job scheduling in Spark

Interface for all trainable schedulers

Base training algorithm class. Each algorithm must implement the abstract method `train_on_rollouts`

Vanilla Policy Gradient