Micron SolidScale – now that’s what I call Big Data!

On May 3rd 2017 Micron introduced its SolidScale enterprise storage solution for early customers to try out ahead of its full commercial launch in Q1 2018. It’s the first ‘boxed’ NVMEoF solution, which sounds complicated, but  (nevertheless) promises a huge leap in the scale, speed and capabilities of storage systems. SolidScale is based on Non-Voatile Memory express over Fabric (NVMeoF), which is a new set of protocols designed to lower the latency and increase bandwidth of server-based flash storage systems. My Figure above shows the quarterly revenues of Micron and Mellanox, which is providing vital chips to SolidScale from 2003 to 2017. I do so to show that there is a role for specialist as well as mainstream chip producers – Micron of Solid State Disk drives and Melanox of computer networking products using InfiniBand and Ethernet technology. Infiniband is a technology we got excited about in my last company way back in 2002, forecasting that 39% of servers would be connected by it by 2005! We’ve had to wait a long time for high-speed interconnects to expand beyond High Performance Computing (HPC) to be relevant in the wider commercial world, but NVMeoF may yet make Infiniband as important as Fibre Channel and Ethernet in the fabric world. The importance of ‘Fabric’ is that flash storage will no longer need to be connected to PCIe buses, allowing vast grids of SSDs to be tied and used together.

Up until now the comparative performance of SSDs with spinning disks has been compromised by the necessity to connect most of them to servers and storage systems using SATA and SAS protocols, which were designed at a time when there was a huge difference between the speed of memory and disk drives. The alternative has been Storage Area Network (SAN) storage, which allows large organisations to share block storage among thousands of users, but is expensive and has much lower response times than DAS. NVMeoF offerings such as SolidScale promise the best of all worlds – liberating SSDs from PCIe bottlenecks and – protocol conversions, while allowing them to be used effectively in huge grids as a pooled, shared storage resource.

SSDs (made up of DRAM and NAND products) now account for the majority of raw storage sales (see my Figure above). The use of SSDs in servers and storage systems has been growing significantly over time, resulting in what was once a high-priced application accelerator offered by specialist start-up vendors becoming a mainstream commercial offering available from all leading systems suppliers, most of whom bought the most promising start-ups.

Micron set up its Enterprise Sales Group some 3 years ago and published its ‘Micron accelerator solutions’ reference architecture, encouraging its OEMs to introduce Remote Direct Memory Access (RDMA) storage. Today you can buy these solutions from server supplier Supermicro. In the case of SolidScale it has been working with the Israeli supplier eXelero. I’m not too concerned that the introduction of SolidScale will see it competing for the enterprise storage market against its vendor customers – it’s a similar move to Western Digital’s acquisition of SanDisk or Seagate’s of Dot Hill. It won’t hinder its OEMs from introducing their own NVMeoF solutions and, in any case, most of its prospects will be cloud service providers, which have a status somewhere between customer and OEM supplier from Micron’s point of view.

SolidScale storage will come in 2 deployment models:

  • ‘hyperconverged’, where compute nodes are the storage nodes and it’s advantage will be in the offloading of processing from the server to the NIC – similar, but faster than the RDMA techniques already in the market,
  • ‘disaggregated storage node’ where the storage nodes connected into a (potentially huge) fabric are the target for management software.

As it’s very new and designed for raw performance Micron’s initial software has only a data services such as protection and failure management. It (and third parties) will no doubt cover richer features over time, although these will have a negative effect on performance. It’s also early days in terms of operating systems (only Linux is available at the moment) and applications (it has run a Microsoft SQL Server database at 11.1GB/s on its 3-node test system and has also been able to run Apache’s noSQL Cassandra database).

SolidScale has some very impressive performance. In particular:

  • Speed – a 3-node test system ran at 10.9m IOPS in Micron’s Austin Texas lab 2 weeks ago.
  • Scalability – a 256 node system was tested in Austin with very little performance tailoff and the company believes it will be possible to stretch to 1k nodes with little tail off in in future.
  • Latency – the company claims that latency for large grids will be similar to DAS, which is important for applications such as fraud detection in the Finance industry.

As a long-term analyst who started his career in 1983 whenever possible I try not to think too much about resolution and capacity issues apart from noting significant improvements like this and remember that it sometimes takes longer than anyone expects for new solutions to become fully adopted (Intel’s Itanium and Infiniband were 2 things which offered the world, but came in late and overpriced). This is the first boxed NVMeoF solution to market; it will be joined by a number of others in the coming year. Although it’ll be some time before significant commercial adoption, it’ll be worthwhile following Micron and SolidScale closely over coming years.