Is the all-flash data centre just a tantalising dream?
The real picture
Flash storage costs can be lowered by using deduplication and compression, but unless you live in a deduplication and compression environment all the time, which means being locked in to a single vendor, then the data has to be rehydrated as it moves beyond the deduping and compression system's boundaries.
The startups bring down the cost of their raw flash by adding deduplication or compression in this way and so claim per-GB costs of $2 to $5, approaching the price of disk. They will claim that, taking into account the data-centre space saved by not having disk drive arrays and the power and cooling savings of not having to spin disks, their all-flash arrays are cheaper.
This idea that deduped flash costs are close to those of disk needs looking at carefully if a rep pokes you in the eye with it. Disk costs vary, with bulk capacity disks approaching $0.60/GB. No doubt if you buy in bulk, disk costs drop even further.
Taking disk drive arrays out of the data centre doesn't save a square millimetre of space
Seagate has just announced 8TB drives and we can surely see 10TB ones coming. You would need to look at the costs of the flash you are being quoted against the cost of the disk capacity that would be replaced to get a like-for-like comparison.
Another grain of salt is needed when considering data centre floor space, as taking disk drive arrays out doesn't save a square millimetre of space; it is still there, just empty, and the air in the data centre still needs conditioning. The savings aren't as obvious or as large as they seem.
If your data centre is a co-location or other hosted facility then replacing disk racks with flash enclosures should deliver a cost saving, but that depends on the facility operator passing on the savings and the percentage may not be what you expect.
Savings here are not black and white – they come in shades of grey.
When data needs to to be taken off the production line and used for a backup and archive resource, where access rates are low and amounts correspondingly huge, then the costs of flash place it at an even greater disadvantage.
There is an emerging form of flash called triple-level cell (TLC) which stores 3-bit values in each cell and is cheaper than mainstream 2-bits-per-cell MLC or the fastest 1-bit-per-cell SLC NAND, however, it is still more costly than bulk disks spinning at 7,200rpm or 5,900rpm and holding 4TB to 6TB. Such TLC flash is also more expensive than archive tape on a cost/GB basis.
Need for speed
Where an archive needs to access data very quickly, flash would be more attractive than disk or tape, but such cases are so rare that some argue there is no real business case for them.
Of course, in theory, such a TLC flash-based archive could be deduplicated to lift its effective capacity, but so far no one is offering such systems. Although TLC flash is used often enough in consumer flash products such as USB memory sticks, its relatively short life of a few hundred write cycles has ruled it out for enterprise use.
That leaves us with the view that all-flash data centres are not feasible at present. They may become feasible if the cost of flash falls to near-parity with nearline and bulk storage disk but there is another problem: the flash foundry capacity to build the stuff just doesn't exist.
In terms of exabytes of capacity, worldwide disk production is vastly higher than that of flash, and with flash fabs costing $7bn to $9bn apiece it is likely to remain so.
This is no small matter. An all-flash data centre would need approximately the same number of TB of storage as current all-disk or hybrid flash/disk data centres.
There is more than just production data to consider: IDC reckons enterprises can have seven or more copies of production data, taking into account all the test and development, protection and enhanced geographic access copies that exist.
The flash foundry operators are paranoid about avoiding loss-making gluts of product, having seen the dire effects of that in the memory industry, with its persistent huge losses and dramatic supplier consolidation. They will be slow to bring new flash fab capacity online.
Stack 'em high
They are working towards increasing flash capacity by increasing wafer density through cell geometry shrinks, and also through building flash chips with stacked layers of cells, so-called 3D NAND.
These in themselves won't allow the flash industry to take on any substantial portion of worldwide disk capacity in the next few years. That requires many new fabs and there is no sign of that happening.
In theoretical terms, then, yes, you could have an all-flash data centre, but it would mean all your non-production data copies and all your protection data stored off-site and online, in the cloud, say. Physically it wouldn't be in your data centre but logically it would.
Therefore, in practical terms, the answer to the question “is an all-flash data centre possible?” is no. Perhaps in five or 10 years' time it may be, but it sure isn't now. ®