Architect Ludwig Mies van der Rohe used the phrase “less is
more” to describe the minimalist aesthetic of his work. While I find minimalism thought-provoking,
I’m a moderate in most things and an unrepentant maximalist when it comes to
dessert.
Technology vendors and consumers relate to tech products
like I relate to dessert. It’s all about
more. Bigger screens, faster CPUs, higher
resolutions, more features. Once the
basics are commoditized, adding features is about the only way for vendors to
differentiate, so that’s what they do.
Data storage is not immune to this condition. Formerly bare-bones storage arrays have been
larded up with features. Snapshots,
tiering, deduplication, compression, clones, replication, scrubbing, analytics
and all the other stuff can be useful. But
there is a cost that has long been hidden.
Modern CPUs are so fast that they can submit a read and have
time to walk to Starbucks for a latte before the storage responds. With traditional hard drives most of this
delay is due to rotational latency. With
SAS SSDs it’s the single-queue inefficiency of SCSI. The storage is the bottleneck. As a result there’s lots of time to run
feature software on the array, and not a lot of pressure for that software to
be efficient.
Now with NVMe, suddenly the storage is orders of magnitude
faster, and the CPU is in much higher demand.
The weight of all those features becomes apparent in the form of lower
performance. Marc Staimer of Dragon
Slayer Consulting refers to this as the “CPU
chokepoint
.” He notes that
you can improve performance by throwing more CPU at the problem, but that there
are seriously diminishing marginal returns.
Storage class memory (SCM) is going to make the situation even worse,
and not even Gordon Moore
can
save us this time. The piper is here
with his hand out.
To maximize the potential of NVMe storage and the bang for
the customer’s buck, everything about the system design, especially the storage
software, must be efficient. Generating
big numbers is important, but if it takes obscene core counts to get there your
product will cost too much.
When Quantum designed our F-Series NVMe storage array
,
efficiency was top of mind. We didn’t
just take some array architecture designed for general-purpose IT and stuff in
NVMe drives. We carefully crafted a
hardware and software system that provides over 24GB/s of throughput in 2U, with
low latency to support multiple streams of 4K and 8K editing and playout
without dropping frames. A single 2U
appliance can do the same job as a full rack or two of legacy storage,
providing space efficiency as well. And F-Series
is cost-efficient, providing a surprising amount of performance per dollar.
Quantum’s most demanding customers have always been in film, where ever-escalating image quality demands huge streaming performance. F-Series was designed for customers like these, who don’t want or need storage arrays with deduplication, cup holders, or power seats. Customers who understand that sometimes less is more. Those who have tried it have been supremely impressed, so I know we got it right. Give us a shout if you want to see how efficient high-performance NVMe storage can be. And while you’re at it cut me a piece of that cake – a big one with lots of frosting.