rdma/rdma_shared_device_a resources allocation

Mellanox / k8s-rdma-shared-dev-plugin

Other

199 stars 35 forks source link

rdma/rdma_shared_device_a resources allocation #92

Open cairong-ai opened 11 months ago

cairong-ai commented 11 months ago

rdmaSharedDevicePlugin: deploy: true resources:

name: rdma_shared_device_a ifNames: [ibs10, ibs11, ibs18, ibs19]

I have deployed the network-operator using the configuration shown above, and I see that on the node it displays: rdma/rdma_shared_device_a: 63. I want to know if these 63 resources are evenly distributing the bandwidth of these IB cards. If a pod only requests 1 rdma/rdma_shared_device_a resource, how much bandwidth can this pod use?

pepesi commented 11 months ago

Based on my understanding, it is merely a device-plugin that exposes the host's InfiniBand (IB) devices to containers. The kubelet can mount them into the containers, allowing the containers to use these devices internally. However, it does not have any mechanisms for bandwidth partitioning or traffic limiting.

adrianchiris commented 11 months ago

the device plugin does not handle any BW allocation (nor am i famiiar with a kernel interface that provides that for the same IB device mounted in different containers), if there is one, let me know :)