HPE’s newly-launched Primera array is a performance beast – and that’s before it gets NVMe over Fabrics and storage class memory.
HPE staffers told Blocks & Files in a product briefing that a 4U 4 controller system could in lab conditions pump out 2.3 million IOPS and 75GB/sec of data with sub-millisecond latency. This approaches high performance computing territory.
Primera achieves this with a pool of memory encircled by four parallelised ASICs per controller. That is a massively-parallel design. Real-world performance may vary but there’s a lot of headroom for that variance.
We learnt the system’s maximum raw disk capacity, in the 4U C650/670 nodes, is 737TB. At 48 drives/node this means 15.36TB drives.
Fibre channel now, NVMe-oF later
Primera will go faster still when it gets system NVMe over Fabrics and storage-class memory Optane drives.
The system hasn’t got NVMe-oF yet because HPE’s customers told it to focus on Fibre Channel first and look at adding NVMe-oF later. They are, we think, risk-averse and regard NVMe-oF as still experimental.
However, NVMe-oF support is already baked into the Primera OS, Ram Gopichandran, HPE’s director for 3PAR portfolio product management, said. A user-installable OS upgrade can be turned on within minutes.
A Primera node can have one or two expansion trays. Gopichandran said there could be NVMe-oF access to these.
HPE is awaiting Intel’s launch of dual-port Optane drives before introducing Optane support. Dual-porting is needed because Primera is designed to be 100 per cent available, and port failure on a single port Optane drive would compromise this feature.
When Intel is ready Primera will get a U.2 Optane drive transplant and, boom, Optane caching/tiering etc. here we come.
Service mentality
Storage interfaces, such as files (NFS or SMB for example), key:value store and object (S3 presumably) are services to be added via the OS update process. Once installed they will run as native services, HPE said.
That means Primera could simultaneously support file, block, object and key:value store access at some stage. (Just to confirm, file services are not included at present.)
The HPErs stressed that the containerised service upgrade process is unique, the system is designed to handle the IO load of NVMe-oF and SCM, and there are lots of exciting things to come.
