Nos is meant to increase GPU utilization and cut down infrastructure and operational costs providing 2 main features:

1. Dynamic GPU Partitioning: you can think of this as a cluster autoscaler for GPUs. Instead of scaling up the number of nodes and GPUs, it dynamically partitions them into smaller “GPU slices”. This ensures that each workload only uses the GPU resources it actually needs, resulting in spare GPU capacity that could be used for other workloads. To partition GPUs, nos leverages Nvidia's MPS and MIG [1,2], finally making them dynamic.

2. Elastic Resource Quota management: it allows to increase the number of Pods running on the cluster by allowing teams (namespaces) to borrow quotas of reserved resources from other teams as long as they are not using them.

Let me know your thoughts on the project in the comments :)

Nos addresses some key challenges of Kubernetes tied to the fact that Kubernetes was not designed to support GPU and AI / machine learning workloads. In Kubernetes, GPUs are managed with [3] Nvidia k8s Device Plugin, which has a few major downsides. First, it requires the allocation of an integer number of GPUs per workload, not allowing workloads to request only fractions of GPU. Second, when enabling GPU shared access either with time-slicing or MIG, the device plugin advertises to Kubernetes a fixed set of GPU resources that do not dynamically adapt to the requests of the Pods at each time.

This often leads to both underutilized GPUs and pending Pods, and/or the cluster admin having to spend a lot of time looking for workarounds to make the best use of GPUs.

For example, consider a company with a k8s cluster with 20 GPUs, where 3 of these GPUs has been reserved to the data science team using Resource Quota objects. In most cases, the workloads of data scientists (notebooks, scripts, etc.) require much less memory/compute resources than those of an entire GPU, yet Kubernetes will force each container to consume an entire GPU. Also, if the team once needs to run a heavy workload, it may want to use as many resources as possible. However, the Resource Quota over their namespace would constrain the team to use at most the 3 GPUs reserved for them, even if the company cluster may be full of unused GPUs!

Instead, with nos the data science team would use nos Dynamic GPU Partitioning to request GPU slices so that many workloads can share the same GPU. Also, Elastic Resource Quotas would allow the team to consume more than the 3 reserved GPUs, borrowing quotas from other teams that are not using them. To recap, the team would be able to launch more Pods and the company would likely need fewer nodes. All this with minimal effort required by the cluster admin, who only has to set up nos.

Let me know what you think of nos, feedback would be very helpful! :) And please leave a star on GitHub if you like this opensource https://github.com/nebuly-ai/nos

Here are some other links that may be useful

- Tutorial on how to use Dynamic GPU Partitioning with Nvidia MIG https://towardsdatascience.com/dynamic-mig-partitioning-in-k...

- Pros and cons of different technologies for sharing resources among workloads in Kubernetes: time-slicing, MIG, and MPS https://docs.nebuly.com/nos/dynamic-gpu-partitioning/partiti...

- Nos documentation https://docs.nebuly.com/nos/overview/

References

[1] Multi-Process Service (MPS) https://docs.nvidia.com/deploy/mps/index.html

[2] Multi-Instance GPU (MIG) https://docs.nvidia.com/datacenter/tesla/mig-user-guide/

[3] NVIDIA k8s Device Plugin https://github.com/NVIDIA/k8s-device-plugin