RLQ Scheduler

---

RLQ: Workload Allocation With Reinforcement Learning in Distributed Queues paper code

Cite us

---

If you use this code, please cite the following paper:

@ARTICLE{StaffolaniRLQ2023,
    author={Staffolani, Alessandro and Darvariu, Victor-Alexandru and Bellavista, Paolo and Musolesi, Mirco},
    journal={IEEE Transactions on Parallel and Distributed Systems}, 
    title={RLQ: Workload Allocation With Reinforcement Learning in Distributed Queues}, 
    year={2023},
    volume={34},
    number={3},
    pages={856-868},
    doi={10.1109/TPDS.2022.3231981}
}

You can contact the authors at: alessandro.staffolani@unibo.it

Deployment

Requirements

---

• Python 3. We utilized Python 3.8
• Kubernetes cluster: used to deploy the system
• A storage class provider. An Open-EBS configuration is provided in kubernetes/open-ebs.
• Local access to the cluster via kubectl cli interface
• A Docker images registry. We utilized Cloud Canister.
• An S3 object storage: used to save the results from training and evaluation. A MinIO deployment using docker is provided in docker/minio.
• MongoDB database: used to save run configuration and results summary. A deployment using docker is provided in docker/mongo/docker-compose.yml.

For our experiments we utilized Kubernetes version 1.2, Docker as Container Runtime Interface (CRI), Calico as Container Network Interface (CNI) and OpenEBS as a storage solution.

Installing local packages

---

The requirements folder contains all the requirements for each component of the RLQ system. If you plan to run only the experiments on the Kubernetes you need to install only the following requirements:

• requirements/base.txt
• requirements/deployer_manager.txt

Otherwise, if you plan to develop locally, you will need to install the requirements for all the components.

Build and Publish docker images

---

Before deploying on Kubernetes it is necessary to build and publish on the docker registry the images of the RLQ components.

The folder scripts contains several scripts to help with this process, before utilizing those scripts you need to create the scripts/.env file with the information to connect to the docker registry (a sample file is provided at scripts/.env.sample).

By launching the scripts/build-and-pull-all.sh script all the required images will be built and published on the registry.

Deploy

Prepare the deployment

---

In the kubernetes/rlq folder it is necessary to create the Kubernetes secret deployments used inside the Kubernetes deployment to interact with the MongoDB, the S3 and the Docker registry instances. The folder already provides for the following sample files:

• kubernetes/rlq/docker-secret.yaml.sample: for the docker registry
• kubernetes/rlq/minio-secret.yaml.sample: for the S3 instance
• kubernetes/rlq/mongo-secret.yaml.sample: for the MongoDB instance

Additionally, the config files need to be changed accordingly to the available settings.

Deployer Manager Config

---

The deployer manager is responsible for deploying RLQ components on the Kubernetes cluster accessible through kubectl. Its config are in the config/deployer_manager.yml file, where it is necessary to change the custom_images property to link to your images. Here the default one:

custom_images:
  - cloud.canister.io:5000/rlq/agent
  - cloud.canister.io:5000/rlq/task_broker
  - cloud.canister.io:5000/rlq/task_generator
  - cloud.canister.io:5000/rlq/worker
  - cloud.canister.io:5000/rlq/trajectory_collector
  - cloud.canister.io:5000/rlq/system_manager

Global Config

---

The config/global.yml file contains a set of configuration available to all components. Unless you change some deployment file these configuration are ready to be used. Please note, here you can change the task classes and worker types configuration.

For the task classes, under the task_classes property it is possible to define the property of a task class. Task classes are Python functions defined in the rlq_scheduler/tasks/tasks.py file

For the worker types, under the worker_classes property it is possible to define the property of a worker type: its replicas, its resources and its costs

Run Config

---

The config/run_config.yml file contains an example of configuration of a single run of the system.

Multi Run Config

---

The config/muti_run_config.yml file allows to configure multiple runs using different parameters for both the agents and the overall system. This configuration is processed by the SystemManager in order to generate one run_config for each parameter provided in the config

Other Config Files

---

The config folder provides additional file configs, one for each component of the RLQ system

Start the deployer manager

---

The deployer manager handles the process of deploying all the RLQ components on the Kubernetes cluster ready to start. In order to launch it, run the following command:

python deploy.py deploy

If you changed any config path or other parameters you can see the full list of arguments of the script using the -h or --help arguments.

When you're done with the deployment you can remove it from Kubernetes by executing: python deploy.py cleanup.

Paper Results

---

The following provides a guide to reproduce the experiments we published in the "RLQ: Workload Allocation With Reinforcement Learning in Distributed Queues" paper.

The experiments are composed of a training and evaluation phase. The training is required to obtain the initial knowledge for the RLQ-LinUCB and RLQ-DoubleDQN agents, while the evaluation is used to compare the performance of our proposed solutions against our baseline algorithms (for additional details about refer the actual paper). In particular, we have performed two main experiments:

1. Synthetic Workload Evaluation: we evaluated using synthetic workload generated by executing simple programs that stress different resources of the workers
2. Real-World Workload Evaluation: we evaluated using Google Borg's traces from 2019

For both evaluations we have tested using three different reward functions: execution-time, execution-cost and waiting-time. In the following the procedure assume your Kubernetes cluster can host at least RLQ environment in parallel (in different namespaces). In this way it is possible to execute in parallel the experiments for the 3 reward functions.

Synthetic Workload Training

---

The deployer manager script has a parameter to launch multiple deploy in parallel. We use this parameter to create the three deployments that are configured in kubernetes/evaluations and config/synthetic-evaluation by running:

python deploy.py deploy --all --folder config/synthetic-evaluation

The config/synthetic-evaluation also contains the multi_run_config.yml used for each experiments.

Note: you might need to modify the custom_images property in the deployer_manager.yml files provided in the configuration folders.

Real-World Workload Training

---

Similarly, to the Synthetic Workload Evaluation in kubernetes/evaluations and config/real-world-evaluation folders is configured RLQ system, thus run:

python deploy.py deploy --all --folder config/real-world-evaluation

Google traces are loaded by the docker images from the data/google-traces folder. However, the google-traces folder contains all the Python scripts and pipeline used to extract those data. Please, refer to the google-traces/README.md documentation for additional details.

Note: you might need to modify the custom_images property in the deployer_manager.yml files provided in the configuration folders.

Evaluation Experiments

---

The steps described so far allows to train the agents, for the final evaluation it is necessary to have a running deployment (see the steps above) and a multi_run_config.yml file with the path to the agents model to load. For simplicity, the generate_eval_config.py script can generate the config file. Before running it is necessary to change the following variables in the file:

RESULT_FOLDER = 'waiting-time'
RESULT_FOLDER_OUTPUT = 'eval-waiting-time'

OUTPUT_FILES = 'config/evaluation/eval_config.yml'

MINIO_BUCKET = 'gtraces1'

AGENT_NAME = 'DoubleDQN'
AGENT_CODE = 'double-dqn'
AGENT_PARAM = ('lr', 'layers')

LOAD_SEEDS_FROM_DB = None  # or None
DB_NAME = 'sb_waiting_time'
DB_COLLECTION = 'waiting-time'

TASK_GENERATOR_SEED = 200
N_TASKS = 13730
N_BOOTSTRAP_TASKS = 0
REWARD_FUNCTION = 'waiting-time'
REWARD_FUNCTION_PARAMETERS = {}
LOAD_MODEL = False
TRAIN = True
ADD_BASELINES = False

Evaluation Plots

---

The folder notebooks contains two jupyter notebook files with the code utilized to generate the plots in the paper. They can be utilize to reproduce the plots. It might be necessary to change the RESULT_FOLDER on the various blocks and to make it point to the save_properties.run_name_prefix property in the multi_run_config.yml used for the experiments