editor - November 11, 2011|
That's a question asked recently by
an analyst in an investment bank. This is of what I said.
PCIe and FC SAN / IP SAN rackmount SSDs can coexist in the same customer
application. Just because a user starts with one doesnt preclude the other.
My March 2010 article
this way to the
petabyte SSD noted there will be 3 distinct types of enterprise SSD in the
future datacenter . Since then I've expanded this model to add more detail and
sub-segments - but overall I think
7 main SSD types
will satisfy all future enterprise needs.
PCIe SSDs make the
locally attached server faster. But SAN SSDs (including
always be needed for environments (legacy command and control business
architectures) where data is shared across more than a threshold number of
servers. Another complication is that a PCIe SSD itself can be a component
inside a rack such as bunch of cards rack (Nextio) or as an internal
cache (NexGen). This is
what I call open / COTS array architecture discussed here-
There are differences in logistics for customers.
rack or a card the replacement unit? (if there's a fault)
sites will have different preferences.
It is technically possible to
design hot replaceable PCIe SSD modules. The concept has been shown to work by
SANRAD but standards on
form factors take years to get established so for now the PCIe SSD market is
based on cards. If I have to leave you with 1 single idea about the
enterprise SSD market its is that no single vendor or technology will be the
ideal solution for the enterprise.
The range of SSD latencies in the pure SSD datacenter of the future
will vary by more than 1,000 to 1.
Design differences mean that the fastest SSD in the user site will
significantly more than the higher tiered SSDs.
will be even more important in the 100% SSD environment. One result from my
model is that it is important for SSD oems to understand which market silos they
If they stray outside (or are unaware of where the market
is heading) their technology and marketing will be inefficient compared to
companies who have optimized their business (by plan or accident) for the model.
Just because Violin sells more SAN rackmounts doesn't mean that FIO
will sell less cards for example.
Selling more PCIe SSDs actually
increases future upstream demand for other types of SSDs. Because an SSD
accelerated server just looks like a faster server.
Hope that helps
Editor's later notes
Decentralization - is an
interesting concept which Fusion-io used in a
S1 filing document to describe their company - "We have pioneered a
next generation storage memory platform for data decentralization."
going into too much more detail you can think about this as follows.
do users use lots of servers?
Because one server isn't fast enough to
do everything they want to do.
But when you have lots of servers - they
often need to access the same data. That's how you get into
architecture - SANs,
- one step removed - the result is longer access times, more complicated
backup, management etc.
you can make a single server run apps 10x or 20x or 100x
faster - by stuffing it with high capacity, low latency SSD - then
you don't need so many
servers (2003 SSD market model).
So in that context - of a medium
size installation - what I said at the start of this article isn't strictly true
- because you could remove the need for a SAN by crunching up all these "no
longer needed" servers and SAN storage into a single box that's easier to
But what I said later in the email is also true. Because
you'd soon find other things you could do to leverage that data. So before too
long you'd be adding the SAN and more rackmount storage back on. And this time
all the new elements would have to run a lot faster to service the higher IOPS
demands of the PCIe SSD accelerated apps hosts.
the SSD Heresies
New Business Case for SSD ASAPs
the Survive and
thrive guide to enterprise SSDs
enterprise SSD market be big enough for all these companies [list] to grow?
|Update - March 2013|
|Within 18 months of publishing this
article - both Fusion-io and Violin had taken additional
product marketing steps to provide more complete offerings to customers thay
had acquired on one side of the server and SAN SSD acceleration
divide - but who
were known to be spending significant budgets in both.|
In the case of
That step was the August 2012 launch of their
- which enabled users to convert a standard server and a bunch of
PCIe SSDs into a
traditional SAN compatible
In the case of Violin...
That step was the March 2013 launch
of their Velocity
PCIe SSD range - marking a long overdue entry into this important segment
of the enterprise SSD market.
However, these 2 SSD companies are
mostly selling into apps environments where they are rarely competing head to
head for the same purchase order.
Why do I say that?
Violin is mostly better suited to
enterprise SSD architectures.
Whereas Fusion-io is better for
enterprise SSD deployments.
Those are big differences in customer
approaches - even if the headline apps functions may look similar to casual
|Oracle users evenly split
between server and SAN when it comes to SSD speedup|
|Editor:- October 11, 2012 - Among other
in a survey of 400 attendees (pdf) which was run by Kaminario at the
recent Oracle OpenWorld
event - it was found that among the 30% of those who had already used flash
SSD acceleration - the use of internal (server based) and external (SAN rack
based) SSDs was split nearly evenly - 48% and 52% respectively.|
|"If there are
hundreds or thousands of application servers, at what point does it stop making
sense adding flash to each server?"|
|Bernhard Behn, Avere Systems - in his blog -
Better than Server Flash? - June 7, 2012.|
Editor:- this was my
re at what point does it stop making sense adding flash to
Because flash in each server means you need less (thousands less)
servers due to of CPU-SSD equivalency.
Most enterprise SSD experts I speak to agree you need both.
More flash on the SAN and more in the servers too.
And this is what happens in user sites which have traditional command
control type apps and who are on their 3rd, 4th or 7th generation SSD
depoloyment. (Its different in democratic server farms like Google.)
For end users who use SSDs for the first time it appears confusing
that they can solve their performance pain point (seemingly) with either server
OR SAN SSD. But thats just an illusion. It may work tactically but isnt
A future point of confusion will be when PCIe SSDs scale out to
connect to hundreds or
thousands of other servers via PCIe. But that architecture will just look
like what we already have in
Infiniband and wont
negate the importance of the SAN.
|"...we address the
toughest problem in front of IT today - where do I put my flash?!!
Flash can go in so many places it makes your head spin, and the more
places you put it, the harder it is to manage it. Fusion IO is telling
you to put it in your server, your storage array companies are saying to add it
to their already overtaxed storage arrays and use tiering software to figure out
your hot data and migrate it there.."
Rods are Cool Again for Big Data - GridIron blog -