open-gpu-server

This repository provides information about the OpenStack instance Quansight and MetroStar are providing to conda-forge and other communities.

Access

The main intent of this service is to provide GPU CI to those conda-forge feedstocks that require it. To do so:

• Feedstocks must have access to the service. Refer to the conda-forge/admin-requests README.
• Maintainers you must read and agree to the Terms of Service. Your username must be listed in access/conda-forge-users.json. Please open a PR to add yourself to the list.

Incidents

If you suspect the server is not operating as expected, please check:

• Status page
• Ongoing incidents

If you think there should be an open incident report but there's none, please open a new issue and tag @Quansight/open-gpu-server so the team can take a look. Thanks!

Base configuration

• Model: AMD EPYC 7352 24-Core Processor
• Architecture: x86_64, 32-bit, 64-bit
• 48 Cores
• ~500 GB Memory
• 6x NVIDIA Tesla V100

Available runners

The server can spin up VMs with the following configurations:

GPU runners

Name	vCPUs	RAM	Disk	GPUs
gpu_tiny	4	2GB	20GB	1x NVIDIA Tesla V100
gpu_medium	4	8GB	50GB	1x NVIDIA Tesla V100
gpu_large	4	12GB	60GB	1x NVIDIA Tesla V100
gpu_xlarge	8	16GB	60GB	1x NVIDIA Tesla V100
gpu_2xlarge	8	32GB	60GB	1x NVIDIA Tesla V100
gpu_4xlarge	8	64GB	60GB	1x NVIDIA Tesla V100

These runners use the ubuntu-2204-nvidia-20230914 image.

CPU runners

Name	vCPUs	RAM	Disk
ci_medium	4	8GB	60GB
ci_large	4	12GB	60GB
ci_xlarge	4	32GB	60GB
ci_2xlarge	8	32GB	60GB
ci_4xlarge	8	64GB	60GB

These runners use the ubuntu-2204-20231018 image.

Software

These runners run ISOs derived from Ubuntu 22.04. Images are built with the instructions provided in the images folder.

Limitations

• Concurrency depends on available resources. Only 4 GPUs can be exposed to the VMs at a time; expect queues. This is not per repository, but a server-wide limitation. See docs/setup.md for details.
• We have not yet implemented a time limit per job. Please be mindful of this and try to keep your jobs as short as possible. This might change in the future.

Support

This service is provided as is, with no guarantees of uptime or support. If you have any questions, please open an issue in this repository and we'll try our best.