A rather tiny bit of storage news this week illustrates the changes taking place as part of the flash revolution, and also where its leading edge lies.
The news is that Fusion-io has submitted proposals for standardised APIs for Atomic writes to the T10 SCSI Storage Interfaces Technical Committee.
Why this is interesting is that it’s is all about the interface between flash memory/storage and some of the most business critical database apps.
Atomic operations are database operations where, for example, there are multiple facets to a single piece of information and you either want all or none of them read/written. Completing only one part is highly undesirable, such as a query for credit and debit in a banking system.
Until now, with spinning disk hard drives, supporting MySQL, for example, because of the possibility of disk drive failures writes took place twice before acknowledgement, as a failsafe. Clearly, such a doubling of operations, is not optimum in terms of efficiency.
What Fusion-io has done is to eliminate that duplication of effort with APIs that build in management of Atomic operations to flash memory.
The flash pioneer claims its Atomic Writes capability provides performance throughput increases of up to 50%, as well as a 4x reduction in latency spikes, compared to running the databases on the same flash memory platform without it.
Gary Orenstein, marketing VP, said: “The background is that many have adopted flash as a direct replacement for the HDD. But Fusion-io believes flash should do more than that and that we should be moving away from the last vestiges of mechanical operations in the datacentre.”
“What we’re looking at are new ways to break new ground that are faster, with fewer instructions,” he added.
Currently these capabilities only come with Fusion-io flash products and are already supported in MariaDB and Percona MySQL distributions but upon T10 standardisation they will be open to all vendors.
Stepping back to take a big-picture view what this also illustrates is the contrast between the extremes of flash implementation in the datacentre.
One the one hand there is this type of work at the leading edge of flash storage use, squeezing ever-greater efficiencies from the interface between flash and the apps/OSs etc that it works with by use of software.
At the other there are the legacy arrays into which flash drives act as a pretty much dumb replacement for the spinning disk HDD.