example_app.py

"""Assumptions:

- Users supply a time estimator func for each node in the DAG.
- Assuming homogeneous instances:
  - Support for choosing # instances
- Support for heterogeneous instances?


Optimization problem:
 - min cost, subject to some constraints?
 - min latency, subject to some constraints?

DAG assumption: chain.  If multiple branches, take into account of parallelism?

Incorporate the notion of region/zone (affects pricing).
Incorporate the notion of per-account egress quota (affects pricing).
"""
import time_estimators

import sky


def make_application():
    """A simple application: train_op -> infer_op."""

    with sky.Dag() as dag:
        # Train.
        train_op = sky.Task(
            'train_op',
            run='python train.py --data_dir=INPUTS[0] --model_dir=OUTPUTS[0]')

        train_op.set_inputs(
            's3://my-imagenet-data',
            # estimated_size_gigabytes=150,
            # estimated_size_gigabytes=1500,
            estimated_size_gigabytes=600,
        )

        # 'CLOUD': saves to the cloud this op ends up executing on.
        train_op.set_outputs('CLOUD://my-model', estimated_size_gigabytes=0.1)

        train_op.set_resources({
            sky.Resources(sky.AWS(), 'p3.2xlarge'),  # 1 V100, EC2.
            sky.Resources(sky.AWS(), 'p3.8xlarge'),  # 4 V100s, EC2.
            # Tuples mean all resources are required.
            sky.Resources(sky.GCP(), 'n1-standard-8', accelerators='tpu-v3-8'),
        })

        train_op.set_time_estimator(time_estimators.resnet50_estimate_runtime)

        # Infer.
        infer_op = sky.Task('infer_op',
                            run='python infer.py --model_dir=INPUTS[0]')

        # Data dependency.
        # FIXME: make the system know this is from train_op's outputs.
        infer_op.set_inputs(train_op.get_outputs(),
                            estimated_size_gigabytes=0.1)

        infer_op.set_resources({
            sky.Resources(sky.AWS(), 'inf1.2xlarge'),
            sky.Resources(sky.AWS(), 'p3.2xlarge'),
            sky.Resources(sky.GCP(), 'n1-standard-4', accelerators='T4'),
            sky.Resources(sky.GCP(), 'n1-standard-8', accelerators='T4'),
        })

        infer_op.set_time_estimator(
            time_estimators.resnet50_infer_estimate_runtime)

        # Chain the tasks (Airflow syntax).
        # The dependency represents data flow.
        train_op >> infer_op

    return dag


dag = make_application()
sky.optimize(dag, minimize=sky.OptimizeTarget.COST)
# sky.optimize(dag, minimize=sky.OptimizeTarget.TIME)