handong32 / cilantro

Source code for OSDI 2023 paper titled "Cilantro - Performance-Aware Resource Allocation for General Objectives via Online Feedback"

Cilantro: Performance-Aware Resource Allocation for General Objectives via Online Feedback

Cilantro is a scheduling framework for performance-aware allocation of resources among competing jobs. Unlike other schedulers, which require jobs to state resource demands, Cilantro applies online-learning to dynamically allocate resources optimizing for the user stated objective. Objectives may be defined on individual utilities reported by jobs, such as latency and throughput SLOs.

For more details, please refer to our OSDI 2023 paper.

Running Cilantro Experiments

Below we link to two examples of running Cilantro on Kubernetes clusters. These experiments were used in our paper.

Microservices experiments

This experiment borrows the hotel-reservation benchmark from Deathstarbench to demonstrate the application of cilantro to an application composed of many microservices. Cilantro can optimize the end-to-end performance of the application by individually controlling the resource allocation of each microservice. See Section 6.2 in the paper for more details.

Please follow experiments/microservices/README.md for instructions on how to run this experiment.

Cluster sharing experiments

This experiment evaluates Cilantro’s multi-tenant policies (Section 4.1.2 in the paper) on a 1000 CPU cluster shared by 20 users. Please refer to Section 6.1 in the paper for more details.

Please follow experiments/cluster_sharing/README.md for instructions on how to run this experiment.

Directory Structure

• ./cilantro contains the core Cilantro code base.
  • ./cilantro/policies contains implementations Cilantro's policies and other baselines.
  • ./cilantro/backends contains different event sources (gRPC, Kubernetes, timer) supported in Cilantro.
  • ./cilantro/learners contains implementations of different online learning algorithms used in Cilantro.
  • ./cilantro/scheduler contains the main scheduler loop.
• ./experiments contains the code for running Cilantro experiments presented in our paper.
• cilantro-workloads is a sister repository that contains the workloads used in our paper. These workloads are already compiled and pushed to Cilantro's ECR repository.

Notes

• If you'd like to build your own image for Cilantro, please refer to ./eks/README.md for instructions on how to create your own image registry on AWS ECR and build/push cilantro images to it.
• ./README_DEV.md contains instructions for developers who want to modify and extend Cilantro.

Citing Cilantro

If you found Cilantro useful, you can cite us:

@inproceedings {288542,
author = {Romil Bhardwaj and Kirthevasan Kandasamy and Asim Biswal and Wenshuo Guo and Benjamin Hindman and Joseph Gonzalez and Michael Jordan and Ion Stoica},
title = {Cilantro: {Performance-Aware} Resource Allocation for General Objectives via Online Feedback},
booktitle = {17th USENIX Symposium on Operating Systems Design and Implementation (OSDI 23)},
year = {2023},
isbn = {978-1-939133-34-2},
address = {Boston, MA},
pages = {623--643},
url = {https://www.usenix.org/conference/osdi23/presentation/bhardwaj},
publisher = {USENIX Association},
month = jul,
}