As part of our overall research cloud infrastructure, we have built a high performance storage cloud service. This is provided for users of the Compute Cloud and the AI Cloud, and is also the basis for the Data Deposit service.
Size and performance
The capacity of the storage cloud is a total of 7,2 PB raw disc space, with failure mechanisms (distributed data to prevent loss) etc.
The cloud storage system is designed to deliver performance that is equivalent or better than traditional local disk drives inside the cloud. Expected storage performance for a node in the cloud:
| Storage metric | Expected performance |
| Write bandwith (for large blocks) | > 300 MB/s |
| Read bandwith (for large blocks) | > 500 MB/s |
| IOPS random write (for small blocks) | > 500 IOPS |
| IOPS random read (for small blocks) | > 500 IOPS |
| Average latency for random writes | < 5ms |
| Avarage latency for random reads | < 5ms |
These numbers are conservative estimates based on our current testing.
Under the hood
We have 720 discs spinning in a five rack setup in our datacenter. Administration of the the cloud is done using Ceph, which is designed for running a distributed storage system delivering high availability and scalability. The current Ceph installation is optimized for high throughput. All nodes in the cloud are connected through a high-speed Ethernet.