Flash storage is on track to achieve parity with high-performance spinning disk on a raw cost per gigabyte basis in as little as a year, according to HPE.
Paul Haverfield, HPE’s data centre and hybrid cloud group chief technologist for APJ, told iTnews it was still probably one to two years "before we get to price parity on a raw $/GB basis, [but] it’s definitely in sight".
“15K RPM disk is already dead and buried today and 10K RPM disk is next,” Haverfield said.
“We’re probably 18 months away before there’s just no economic use case to buy 10K spinning disks.
“What that does is it moves the focus onto 7K RPM bulk capacity drive that’s 10, 12, 14TB and that becomes next in sight for QLC NAND flash technology.”
To date, enterprise adoption of all-flash and hybrid arrays has been largely predicated on using “compaction” technologies like compression and deduplication to fit more data into the raw data space on the array.
This is often expressed as an “effective” or “usable” $/GB, but achieving it is contingent on the enterprise customer compressing data so they can store more in the available raw space.
Back in 2015, when iTnews first examined the all-flash space, one needed to achieve a compaction ratio of 4:1 for the usable $/GB to be roughly in line with the raw $/GB cost of spinning disk.
However, in two short years - and despite recent shortages of NAND flash drives - the ratio has shrunk to 2:1, according to Haverfield.
“Still to this day to make flash storage price equivalent to high performance spinning disk on a cost per gigabyte basis you typically need a compaction ratio - includes compression, dedupe, etc - of at least 2:1 to make flash price equivalent on a $/GB basis compared to 10K RPM spinning disk,” he said.
But give it a year - two years maximum - and flash storage is anticipated to cost the same as high-performance spinning disk on a raw - not usable - $/GB basis, he said.
Once there is parity in raw $/GB costs, Haverfield believes NAND flash SSDs will be a more attractive option than spinning disk, since costs can be driven even lower by pursuing some form of data compaction.
“If you can easily get 2:1 compaction across your online workloads why wouldn’t you?” he said.
“Alternatively, you can buy the same amount of raw capacity and just do double the number of snapshotting virtual copies [for backups]. You can have more protection up your sleeve.”
To dedupe or not to dedupe? It might not be your choice
That is, of course, assuming the flash array allows compaction technologies to be switched on or off - that is, they aren’t simply on by default.
“Today you don’t have a choice [on some all-flash arrays],” Gartner research vice president Arun Chandrasekaran said.
“These technologies are in-line which - in summary - means you can’t turn them off. They’re automatically on. It’s part of the base software license.”
Back in 2015, some enterprise customers were wary in particular of de-duplication, preferring not to use it and simply absorb the higher $/GB cost of flash.
Given that isn’t always a choice, adoption of dedupe and other compaction technologies “has accelerated”, Chandrasekaran said.
However, even when they had a choice, Chandrasekaran said in his experience “very few customers turn them off".
Haverfield said while there were “valid uses cases for both” configurations, granular control of dedupe and compaction was arguably still more useful.
“A lot of customers migrating all-disk arrays into all-flash - which all customers will do over the next 2-3 years, there’s no doubt whatsoever about that - a lot find their data doesn’t natively deduplicate very well,” he said.
“With a solution which is attempting to dedupe all the time, that consumes CPU power and memory in the storage system.
“Every architecture with dedupe always-on is very limited by the amount of cache memory the hardware has and the deduplication data structures or the metadata to support dedupe and how it’s stored in that device’s memory. So dedupe does affect architectural scalability.”
Haverfield also claimed compression was becoming more common at the application layer, meaning there may be limited gains to be achieved from applying it to the workload’s storage.
“There’s a growing trend where a lot of transaction-based applications are supporting compression services at the application or database level,” he said.
“Those applications just do not have any further dedupe or compression available at the storage backend … so it just becomes overhead and waste if you try to deduplicate. It’s just not possible.
“If Oracle is already doing Oracle Advanced Compression at the table level, you might get another 1-5 percent compression in your storage array, but is it worth it given the amount of CPU and latency it might add back on reads? It’s probably not worth it.”