The last white space in the NAND flash market segments (see Footnotes for complete list) is the very low latency (microsecond to nanosecond) Server SAN segment. Short distances and highly efficient protocols allow very rapid and very high bandwidth communication between flash storage units very closely connected to server processors as an eXtension of memory - hence 'Xflash.
The two major vendors who have intimated intentions in this area are Fusion-IO with its ION product, and EMC with its recent purchase of ScaleIO.
There are two major market sub-segments for Xflash:
- The High Performance Application Segment: The market for this is initially for very large system databases that mandate very low write latency and very low interconnect speeds to meet the large-scale system objectives.
- The HPC/Big Data Analytics Segment: In this market flash storage is used mainly as an intermediate persistent storage of data.
The larger is the High Performance Application Segment. Figure 1 shows the fundamental integration of event streams into real-time processing of both transactional and data analytic data, providing near real-time feedback and near real-time automated modification of the transactional systems. Xflash is required to achieve this level of performance.
Source: Designing Systems and Infrastructure in the Big Data IO Centric Era, Wikibon 2012
Flash-only arrays operate at about 1 millisecond (10-3) read or write elapsed time. A large part of the remaining latency is storage network and SCSI protocol overhead. By reducing the network protocol overhead and closely positioning the servers, XFlash allow sub microsecond (10-7) IO latency and billions of IOPS.
A fundamental technology of the IO stack that has to be eliminated is the “n-phase commit structure of SCSI”. This is needed to guarantee data integrity in traditional very slow access to disk environments and over traditional SAN fabrics. The method used to elimination this overhead is atomic writes, using NAND flash as an eXtension of memory, or Xflash. Atomic writes were first demonstrated by Fusion-io, and an Atomic Writes API has been contributed for standardization to the T10 SCSI Storage Interfaces Technical Committee by HP and Fusion-io. It is now in use in mainstream MySQL databases such as MariaDB 5.5.31 (Beta) and Percona Server 5.5.31. The de facto interconnect technology for Xflash is RDMA over InfiniBand.
Wikibon believes that Xflash will be the only NVM technology of importance for enterprise computing for at least the next five years, and probably much longer. Although there has been much discussion about replacing flash with some other type of NVRAM technology (e.g., MRAM, ReRAM or PCM), the very low elapsed time for writing out to the flash controller and receiving an acknowledgement puts in question the future of other non-volatile memory (NVM) contenders other than for niche usage. Unless a new NVM technology is adopted in consumer products in volume, the continued investment in flash technologies driven by consumer demand will make volume adoption of any NVM alternative very unlikely, and make their adoption in enterprise computing economically impossible.
Figure 2 shows the Wikibon projected relative size of the storage capacity and spend of different storage technologies. The Xflash market is projected by Wikibon to grow to be 20% of the total storage market by 2020.
Source: Flash and Hyperscale Changing Database and System Design Forever, Wikibon 2013
Action Item: CEOs and CIOs should understand the business potential of Xflash systems to provide low-cost high-function alternatives to traditional IT systems, and ensure that strategic plans are in place to adopt such technologies as quickly as possible.
Footnotes: The Wikibon NAND flash market segments are:
- Hyperscale – e.g., Fusion-io ioScale
- Unified Hybrid Storage – e.g., EMC VNX, Tinrí (2-20 milliseconds)
- Tier-1 Hybrid Storage – e.g., HDS VSP & HP 3PAR (2-20 milliseconds)
- Flash-only Array – e.g., EMC XtremIO, IBM Texas Memory (~1 milliseconds)
- Xflash – e.g., EMC ScaleIO, Fusion-io ION, (<100 microseconds)
