ZTP Stands for Zero Touch Provisioning: Automated OpenShift deployment that setup registry mirror, nodes,etc out of a configuration yaml.
The architecture was described as Openshift Container Agnostic TOPology Integrated Chassis. The architecture was based on a 3+1 chassis which had everything needed for the automation with nics for internal/external networks.
It’s an operator that builds custom kernel modules to provide support for specific hardware in the systems.
About to get 1.0 released with some features coming for 1.1:
What is a DPU? :
The DPU is can be used in several ways:
The Current design is using Shared-OVN with 2 clusters:
The cluster will communicate between them and setup a shared OVN network for offloading POD’s traffic
Links:
DMA what?
RDMA is a kernel feature to share buffers between hosts. The general idea is setting up a peer-to-peer DMA between NIC and transferring / sharing buffers for GPU workloads.
Starting in RHCOS 9 :
In the old way, each VM we create has a PCIE pass-through to GPU. which means that only one application / service can run on one GPU.
The Nvidia GPU operator is now extended and supports vGPU - taking one physical GPU and sharing to multiple instances.
The next evolution of GPU is using MIG ( multi instance GPU ) which partition a GPU based on Ampere architecture ( A100 + ) to different hardware slices which is more secure /isolated and better performance.
Links: