Pavilion has lots of ambitions

Pavilion Data Systems, emerging leader in high-performance unified storage, delivered a super session during the recent digital IT Press Tour.

Started historically as a block storage array, full flash, with impressive characteristics, the Pavilion array named today Hyperparallel Flash Array (HFA) shows really high performance for its density. It's a real challenge, the mission is clear and ambitious: offer the DAS performance with a shared device benefits.

To position itself, Pavilion promotes HFA as the leader of a 3rd wave of storage with first wave being HDD and second wave with Flash/SSD. This timeline is very device centric and by device I mean disk itself and I regret the non consideration of connectivity and architecture as FC/SAN was a key milestone for the industry especially as HFA accepts NVMe SSDs but above all NVMe-oF. In a very rapid shortcut, we can assimilate NVMe-oF as the transport of NVMe like FC did/does for SCSI commands.

The HFA is a 4U chassis with 20 controllers - 10 dual controller cards - exposing iSCSI, FC and NVMe-oF (TCP and RDMA i/e IB and RoCE v2) but also NFSv3 & v4 and S3 with an embedded MinIO engine. This is what the industry named Unified Storage and I use several times the acronym BFO for Block, Files and Object to qualify the offering.

Internally the system receives 72 NVMe U.2 SSD organized in 4 groups of 18 drives. Drives and controllers are connected via a redundant 6.4Tb/s PCIe switch and each drive is dual ported to facilitate failover in case of controllers failures. A controller is not built with FPGAs nor ASICs as it is essentially an optimized and enriched Linux running on Xeon CPUs with a bunch of DRAM offering 2 networks ports thus a dual controller board delivers 4 x 100GbE/IB ports. The array implements a cacheless model with a shared PCIe memory for metadata that allows each controller to immediately serves consistent data to hosts.

The total capacity of a full HFA reaches 1.1PB raw built from 72 x 16TB, exactly 15.36TB SSDs, but the reality is different as a group of disks is internally protected by a RAID 6 implementation within each group. Therefore the hardware overhead is 12.5% (=18/16) and if a spare is enabled per group it means 13.3% overhead so pretty similar. In that case a group is divided with 15 data drives + 2 parities + 1 spare. Globally usable capacity dropped to 920TB with strong protection i.e dual parity plus spare. For the storage space efficiency with any of data processing on that, HFA can deliver 83.3% of usable storage on raw.

In term of performance, the array is pretty impressive with 120GB/s, 20M IOPS and 100 micro seconds of latency. The following slide illustrated perfectly the capability of the product.

In term of data services beyond access methods, Pavilion OS provides:

• Thin Provisioning
• RAID 6 and snapshots
• Encryption-at-rest
• and Multipathing

About access methods exposed by the HFA, as said, 3 different ones are available but a controller is able to expose only one. A protocol namespace belongs to only one controller today being associated with one disk group so up to 18 SSDs. Later this year, namespaces will be able to span controllers but with still the limit of namespace per controller. And we expect that namespace will also span chassis in the future. As namespaces are completely separated, it's not possible today to expose a namespace via multiple protocols especially between file-base protocols like NFS and object-based like S3. Also, SMB is not identified as a need from prospects and users. It confirms that HFA targets specific use cases where performance is a key requirement.

Reduction with compression and async replication across chassis is on the roadmap to make the array hyper resilient and resistant to complete array downtime or site failures or just inaccessibility.

On the business side, Pavilion is 100% channel and leverages key partners to sell its product. The company is a Silver Business Partner with IBM, has relationship with HPE in the field and Dell.

Pavilion is already deployed in one of the largest NVMe-oF configurations at TACC, the Texas Advanced Computing Center, where multiple HFAs are coupled with IBM Spectrum Scale in a shared array model. The company is already adopted by several high demanding environments where access data requires a high SLA.

A competitor with similar ideas, Vexata, hit a wall a few quarters ago due to too long sales cycle that finally killed them. Close to the bankruptcy, StorCentric was able to takeover the company and the engineering team at a very attractive price.

It seems that Pavilion team understands pretty well market fluctuations and impacts. Definitely a storage player to keep under the radar.