Run:ai
A Kubernetes-based GPU cluster scheduling platform that provides queues, quotas, priorities, and preemption mechanisms, turning GPUs from exclusively held devices into shared, schedulable pooled resources. It was later acquired by NVIDIA and open-sourced.
Timeline (last 90 days)
KAI Scheduler originated from an internal scheduler project at Run:ai and uses gang scheduling, meaning all workers of a distributed training job are either scheduled at the same time or all wait.
NVIDIA open-sourced the entire Run:ai GPU orchestration platform under the Apache 2.0 license.
Run:ai was initially known for helping enterprises efficiently schedule GPU resources on Kubernetes, and after being acquired by NVIDIA its positioning shifted toward being the hub of the AI factory.
Run:ai introduces queues, quotas, priorities, and preemption mechanisms on top of Kubernetes, turning GPUs from exclusive devices into pooled resources that can be shared and scheduled by multiple teams.
After open-sourcing Run:ai, scheduling policies, resource quota logic, and priority algorithms can all be inspected and modified, improving auditability and customizability.
Currently the officially disclosed information is concentrated at the level of open-sourcing the complete platform; the specific code repository structure, license type, and community governance model have not yet been elaborated in detail.
Run:ai's core scheduling components may be released under Apache 2.0 or a similar permissive license, while the parts related to enterprise-grade support may retain a commercial option, which is an open-core model.
When NVIDIA opens up the scheduling layer for free, the value proposition of third-party GPU schedulers needs to be rethought.
The deep coupling of Run:ai with NVIDIA hardware and the DSX platform means that the freedom brought by the open-source code will noticeably diminish when stepping outside the NVIDIA ecosystem.
NVIDIA announced that it will fully open-source its Run:ai platform, as part of the NVIDIA DSX AI Factory platform strategy.
1 more timeline events
All Facts (10)
KAI Scheduler originated from an internal scheduler project at Run:ai and uses gang scheduling, meaning all workers of a distributed training job are either scheduled at the same time or all wait.
50%UnverifiedRun:ai was initially known for helping enterprises efficiently schedule GPU resources on Kubernetes, and after being acquired by NVIDIA its positioning shifted toward being the hub of the AI factory.
50%UnverifiedRun:ai introduces queues, quotas, priorities, and preemption mechanisms on top of Kubernetes, turning GPUs from exclusive devices into pooled resources that can be shared and scheduled by multiple teams.
50%UnverifiedAfter open-sourcing Run:ai, scheduling policies, resource quota logic, and priority algorithms can all be inspected and modified, improving auditability and customizability.
50%UnverifiedThe deep coupling of Run:ai with NVIDIA hardware and the DSX platform means that the freedom brought by the open-source code will noticeably diminish when stepping outside the NVIDIA ecosystem.
50%UnverifiedWhen NVIDIA opens up the scheduling layer for free, the value proposition of third-party GPU schedulers needs to be rethought.
50%UnverifiedSchedulers such as Volcano (a CNCF project) and Run:ai provide features including GPU sharing, preemptive scheduling, and elastic scaling.
50%UnverifiedNVIDIA open-sourced the entire Run:ai GPU orchestration platform under the Apache 2.0 license.
50%UnverifiedRun:ai's core scheduling components may be released under Apache 2.0 or a similar permissive license, while the parts related to enterprise-grade support may retain a commercial option, which is an open-core model.
50%UnverifiedNVIDIA announced that it will fully open-source its Run:ai platform, as part of the NVIDIA DSX AI Factory platform strategy.
50%