Sergey Nivens - Fotolia

Datrium gets Flash End-to-End in its server/shared hybrid DVX NVMe

Taking aim at hyper-converged, Datrium DVX adds flash as bulk storage, while NVMe at the server handles hot data with the scale needed by enterprise customers in mind

NVMe-focused storage hardware maker Datrium has upgraded its DVX products to include flash in host and main storage nodes.

With so-called Flash End-to-End, Datrium gives customers the option to populate its shared bulk storage nodes with all-flash storage.

This contrasts with Datrium’s existing hardware offering, which sees NVMe flash cards in host servers, also known as DVX Compute Nodes, while bulk storage is handled by 7,200 rpm SATA spinning disk.

Its disk-equipped DVX Data Nodes have a maximum raw capacity of 29TB (83TB after data reduction), while flash data nodes comprise 16TB of flash (up to 48TB effective).

The upgrade means that Datrium will allow customers that suffer failure at host flash to operate at the speeds of shared flash storage. Datrium marketing vice-president Craig Nunes said Flash End-to-End gives customers two times the bandwidth compared with the spinning disk version.

Datrium is one of a wave of flash storage makers that is focused on providing NVMe-based products. NVMe is a subset of PCIe and comes in a card form factor.

As a protocol, NVMe offers hugely increased input/output per second (IOPS) and lower latencies than existing flash products that use spinning disk-era connectivity methods and the SCSI protocol.

By doing away with SCSI, NVMe potentially boosts flash performance exponentially. That can be achieved by simply slotting it into servers, but currently a key stumbling block to achieving NVMe’s potential in a shared storage environment is the ability to handle controller functionality at speeds that don’t bottleneck input/output (I/O).

Datrium’s answer to this conundrum is to place NVMe cards and controller functionality in host servers, which is an approach somewhat similar to that of E8. Here it claims a performance premium of 2x to 4x over SCSI-connected flash, while bulk storage resides on a shared array-like device.

Read more about NVMe

“We split the storage controller from the durable data [bulk storage] and put the software functions – data deduplication, erasure coding, encryption – on the server where CPU resources are abundant and cost efficient. Meanwhile, the durable data is where the authoritative copy of the data resides,” said Nunes.

Datrium does its best to make all reads local from NVMe on the server, but writes are routed to the shared storage Data Node.

“Servers are effectively stateless, which eliminates a lot of network traffic, noisy neighbour problems, and that means we can mix workloads freely,” said Nunes.

“In read-heavy situations, we get the full benefits of NVMe. When writes occur, they are mirrored to data nodes and acknowledged back, similarly to all-flash,” he added.

Datrium compares its approach favourably in contrast to hyper-converged infrastructure, which works on scale-out clusters of server and storage in single nodes.

“Every HCI node is a storage node, and that is an inherent issue at enterprise-scale deployments,” said Nunes. ... ... ... ... .... .... .... .... .... .... ...

Datrium aims at enterprise-scale deployments in, for example, virtual desktops, databases and the internet of things (IoT). ... ... ... ... ... ... .... ... ... ...

Read more on Hyper-converged infrastructure