Awesome
Clockwork
A multi-tenant managed inference server, backed by a modified version of TVM. Read about Clockwork in our OSDI 2020 Preprint
This README file describes the pre-requisites and steps required to build and run Clockwork. If you follow these steps but encounter errors, please e-mail the mailing list.
Clockwork is not feature complete, but we welcome contributions from others!
Mailing List: clockwork-users@googlegroups.com
Resources
Other Repositories
The following other repositories are relevant and will be referenced here and there.
clockwork-resultscontains experiment scripts and documentation for reproducing results from the OSDI 2020 Clockwork paper.clockwork-modelzoo-voltacontains pre-compiled models that can be used for experimentationazure-functionscontains workload traces from Microsoft Azure that can be used for experimentationazure-functions(deprecated) contains the "preview" traces from Microsoft Azure. This repository is only available internally. Credentials will be provided to OSDI 2020 evaluators.
Getting Started
The following pages step through the things required to build and run Clockwork
- Installation Pre-Requisites
- Building Clockwork
- Environment Setup
- Clockwork Configuration
- Running Clockwork for the first time
Next Steps
- Clockwork Workflow An overview of Clockwork's current workflow
- Customizing Your Environment Tweaks needed if you have different machines and GPUs
- Running Without GPUs Instructions for running without GPUs
Additional Information
- Telemetry Description of telemetry logged by Clockwork
- Troubleshooting Guide Common error messages
- Experiment documentation in the
clockwork-resultsrepository. - Workloads Available client workloads
- Controller Controller options
Contacts
Mailing List
clockwork-users@googlegroups.com
People
Arpan Gujarati, Reza Karimi, Safya Alzayat, Wei Hao, Antoine Kaufmann, Ymir Vigfusson, Jonathan Mace
Organizations
Max Planck Institute for Software Systems
Emory University