 |
VOOZH | about |
|
VOOZH | about |
We’re so glad you’re here. You can expect all the best TNS content to arrive Monday through Friday to keep you on top of the news and at the top of your game.
Check your inbox for a confirmation email where you can adjust your preferences and even join additional groups.
Follow TNS on your favorite social media networks.
Become a TNS follower on LinkedIn.
Check out the latest featured and trending stories while you wait for your first TNS newsletter.
At KubeCon Europe, NVIDIA announced today that it is open sourcing KAI Scheduler, a GPU-centric Kubernetes scheduler that was originally developed by Run:ai, which NVIDIA acquired last year. Available under the Apache 2.0 license, KAI Scheduler helps its users optimize GPU resource allocations for AI and machine learning workloads in GPU clusters.
NVIDIA argues that traditional resource schedulers are ill-suited for managing AI workloads because GPU demand can fluctuate quite a lot, with bursty inference workloads and sustained model training runs that can extend over days.
KAI Scheduler promises to give these teams a better tool for managing these workloads by, among other things, dynamically adjusting quotas and limits in real time, while also offering a variety of scheduling strategies — gang scheduling, hierarchical queuing, bin-packing, spreading and GPU sharing — to avoid long wait times for access to GPUs.
Sharing GPUs looks like it will be an especially useful feature here. This allows multiple pods to utilize the same GPU, for example. It’s worth noting that NVIDIA already offers a tool called GPU Operator, a Kubernetes framework for provisioning GPUs, which also includes a GPU time-slicing feature.
GPU Operator, however, is very much focused on working with NVIDIA hardware and large clusters (including NVIDIA’s own DGX racks), while KAI Scheduler is more vendor-agnostic and also supports AI workloads on CPUs.
KAI Scheduler’s approach, other than GPU sharing, focuses on the individual GPUs and the memory available to them. What developers can reserve here is a share of that memory. There is no memory isolation, though.
By default, KAI Scheduler integrates with popular AI tools and cloud native frameworks like Kubeflow’s Training Operator, Ray and Argo.
The code and documentation for KAI Scheduler is now available on GitHub. Quite a few other parts of Run:ai are already open source, too, including the somewhat related Genv GPU environment and cluster management tools.
