Volumez introduces a SaaS based orchestration running in AWS that controls the deployment of block storage volume running on-premise or in clouds such as AWS, Azure or OCI. Just an agent, named a connector, must be installed in each media node which hosts NVMe SSDs. This connector talks with the SaaS service to receive information about block storage volume to create with local standard Linux commands such as lvm or mdadm. That connector runs in user space and is obviously out of the data path serving only for configuration purposes. It means that if it fails the current running or declared volumes are not impacted.
Two configurations could be considered:
- Converged one with application running on the same node with local medias. Multiple applications can run on the node.
- Disaggregated one with applications running on separate nodes connected to a layer of media nodes exposing block volumes over the network. Each volume is single attached to an application node.
The solution comes with 4 policies that can be copied and modified. There is no cluster, volumes are attached and visible to only one node, normal for converged configuration, and all redundancy is made by the linux tools mentioned above.
We understand that the intelligence comes for the inventory made by the engine that profiles all SSDs collecting characteristics to allocate correctly applications based on available IOPS and capacity. But it appears that there is a current limitation with only one SSD per media node to better control usage, performance requirements and QoS. In that case, it means that configuration requires very large horizontal setup leveraging wide striping with or without redundancy depending of the role of the volume between ephemeral or persistent desires. This is transparent for users and it costs the same thing so this large configuration doesn't introduce any issue. I will cover how DIaaS can be used for AI oriented workloads.
0 commentaires:
Post a Comment