They jumped into the wagon to deliver a scalable disaggregated block storage fueled by end-to-end NVMe. The architecture reminds me what some vendors are doing as they leverage the any-to-any connection between access nodes and storage nodes. They’re using Intel SPDK so everything is running in user space.
The architecture has 2 layers, the first is the back-end storage where the storage node with physical NVMe devices reside exposing native block interfaces via InfiniBand and RoCE connectivity. The second is the controller layer or access nodes, connected to the back-end that expose logical storage entities to clients machines, connected with RDMA/TCP. The beauty of this model is the independent scalability between the controller and storage nodes able to offer a multipath for clients. The technique used here is based on sharding with volumes controlled by the access nodes working in a stateless mode.
In terms of data protection, the development team implements replication within the logical array.
The engineering team develops a write anywhere data placement layer based on a log-structured model and everything is parallelized. Thus they optimize how blocks are written on final NVMe medias having a positive effect on the endurance. The goal is to deliver a sub-millisecond latency system with a throughput of 9GB/s per controller or access node and of course enable NVMe-oF. The first release planned for February 2023 will scale to a few 10s of PB but the ultimate goal is multiple of 100s of PB.
This project is very ambitious and plays against famous names such Kioxia Kumoscale, Lightbits Labs or ones from usual suspects. And we also remember stories that hit the wall like Excelero, acquired for a penny by Nvidia, E8 Storage, swallowed by AWS, or Exten Technologies disappeared inside OVHcloud. They also have in mind some other pioneers we met in the past like Apeiron or Mangstor morphed into Exten. More info soon as the product should be launched in February 2023.
0 commentaires:
Post a Comment