| leading the way to the
new storage frontier
Fusion-io's CEO David Flynn - talks SSD - with StorageSearch.com
|Editor:- July 23, 2010 - I've been
2007 - and I
thought I already knew a lot about the company. |
But I was pleasantly
surprised to learn a lot of new things - in a talk I had with their CEO David Flynn,
yesterday - which you might be interested to learn about too.
decouples storage software from 30 years of data spin
|Editor:- Fusion-io recently
what it calls a "Virtual Storage Layer" which provides a set of
enhanced programmatic interfaces for optimizing operations which use its
ioMemory family of SSDs.|
I had a briefing with Fusion-io's CEO David Flynn ostensibly to
fill me in on more details about the new software tools - but as you can imagine
- once we started talking about what's happening in the SSD market - and what
we think might happen - we covered a wide range of topics - enough to fill
several future articles.
The thinking behind the VSL is to provide
software tools which enable developers to communicate in the new language of
directly accessible flash SSD memory in a way which breaks away from the
cumbersome restrictions and limitations of 30 year old software models which
are layered on legacy hard disk sectors.
One obvious advantage of this
method is performance. By using the VSL programmers have direct access to the
large block writes and random reads which are characteristic of native flash
storage. This visibility eliminates the risk - which can arise when working
via an artificial HDD based I/O software model - that arms-length software can
sometimes result in worsening the performance with an SSD instead of improving
The new SSD API can also enable developers to surgically remove
operations which were originally created to ensure good operation with HDDs but
which you don't need to do with Fusion-io's flash drives.
David Flynn cited an example of a write process in MySQL - in which
2 successive writes take place - as a defensive mechanism to ensure that data
has been written to the hard drive and doesn't get lost if the power goes down.
The 2nd write isn't needed (as an integrity operation) for their
flash SSD - because it will guarantee the data is written even if the power
supply suddenly turns off. In fact doing the 2nd identical write to the same
flash block could in most flash SSDs delay the next microscopic database
The benefit of giving programmers an alternative
programming model - is that functions like this - which were originally
developed for rotating disk storage - can be removed or modified.
performance improvements can accrue not just from accelerating the same
old operations faster - but by rewriting kernel code to improve its data
form factor fit with the new nature of flash memory.
a subject which we spoke a lot about - and I'll return to some of what we
discussed in future articles.
Most flash SSD designs today were
designed to interface through a traditional disk storage interface - like
SAS. That means they
have an internal
controller which translates the host I/O requests which come through the "disk"
interface and translate these into block R/W requests for the flash memory - as
well as doing a lot of other
as we've discussed in these pages before.
Most of the
microprocessors (with a few rare exceptions) which are used as SSD
controllers don't have end to end
checks built into the datapaths of the CPU itself. Consequently they are
vulnerable to cosmic radiation events which can destroy the integrity of the
remapping tables and make the data
David said that in his past career experience working in
supercomputer environments - with big hard disk populations - those
statistically rare events could cause systems to crash on a daily basis -
making it necessary to investigate the weak design points of every single type
of device in those networks until the culprits could be found and modified or
risk exists in large
Partly for this reason he said Fusion-io's product
architecture doesn't include a processor on the PCIe card to perform the
remapping functions. Instead remapping is done by the server host (and server
makers have learned from
experience to make sure their datapaths are end to end corrected.) David
said there's a virtuous triangle which comes from this approach.
Removing the need for a separate SSD controller reduces hardware count, and
reduces cost while
need a fast enough multi-core host processor to handle this task- but then
that's why you're using this type of SSD in the 1st place - because you have CPU
performance which wants to get to work on the data.
supports Linux, Windows 64, FreeBSD, Solaris and VMware.
talking to David I could see that if developers started writing applets to tweak
the speed and reliability features of Fusion-io's SSDs - that would make users
reluctant to look at any of the 100 or so other
PCIe SSD companies
which could be in the market next year.
|| David agreed that
customer stickiness of this type would not be something he would be unhappy
|Here's a sprinkling of the topics
discussed when Fusion-io's CEO - David Flynn spoke to the mouse site.|
that people have turned to SSDs for server acceleration because CPU clock
rates stopped accelerating and
disk spin rates stalled - but David Flynn suprised me by saying that in his
view the DRAM market had run out of steam too.|
At the chip level the
storage capacity of flash is much denser than DRAM.
But because of
the lower power consumption of flash you can pack much more flash into a
than the chip density comparisons would suggest.
He said that using
today's technology you can get 200x more capacity using flash than RAM
in a rackmount system. He said that
installations are technically feasible now.
|David Flynn said that in the
near future datacenters would be regarded as drivers of the knowledge economy
in the same way as factories were in the industrial age.|
|flaky SSDs and flash
burnout in RAID|
|David Flynn said the situation
regarding the misapplication of consumer grade SSD technology to enterprise
applications was a real problem for the reputation of the SSD market.|
said one organization (which he named) had installed RAID systems using
Intel SSDs in a high
About half the SSDs had "burned out"
after a year.
Worse than that - when the customer investigated more
closely they found that some SSDs had failed in a way which had not been
detected by the RAID controllers.
That meant the data was trashed.
Flynn wryly commented that there was a lot more to marketing enterprise SSDs
than adding an "e" to a consumer technology SSD brand (and
redesignating it an "enterprise" product).
He said Intel has
a very strong brand
in the computer market and he doubted it would be dented - even by as many as
4 SSD design recalls
about SSDs is a serious issue in the industry.
some competing PCIe SSD architectures /
the SSD Heresies|
|David Flynn shared his views
about other oems - rushing to get into the PCIe SSD market.|
products had been little more than cards which had a
RAID controller and
bunch of SATA SSDs
repackaged on a card. (For example
- unveiled at CeBIT in March
said not only was performance terrible - because they were doing all that hard
disk interface "rubbish" between the PCIe bus and the flash memory -
but the products did not have end to end error correction. Any big installed
base of such SSDs ran serious risks of uncorrectable random data corruption.
said Fusion-io's attention to the subject of data integrity was one of the
things which was appreciated and well understood by their server oem partners
said that some other new companies coming into the PCIe SSD market with native
PCIe designs - such as LSI
(working with Seagate)
might find it difficult to solve subtle reliability problems - because the
different parts and technologies (which include controllers from
SandForce) would be
coming from 4 (or more) different places.
|David Flynn talked more about
this subject than anything else. |
Mostly it was related to SSDs and is
reported elsewhere on this page.
But he also said that it was a bigger
problem for hard disk based systems than most users fully appreciated - because
of inadequate strength error correction designed into industry standard hard
He said the only practical way to prevent data
corruption in HDD based datacenters was to overlay additional levels of error
checking and data healing outside the arrays. This harmed performance - but was
He observed that many mid range
RAID vendors approach big
end user sites
dislodge vendors like
EMC with lower cost
storage. But one of the added values in these big iron storage systems is the
enterprise wide data integrity cushion which mid scale vendors can't easily
penetrate or replicate.
|David Flynn talked about the
virtues of stripping out layers of unnecesary junk in the design of their SSD -
which would only be there to emulate hard disk protocols.|
He said this
was one of the aspects of the company which had appealed to
Steve Wozniak - who is
now Fusion-io's Chief Scientist.
Flynn related one of the
many legendary tales
told about Wozniak's design of the Apple Computer - the world's first high
volume desktop PC - when Wozniak replaced a whole bunch of chips which
traditionally would have been needed to interface to a floppy drive - and
instead used sofware.
Not only did that save cost - but by using the
6502 processor to calculate the momentun of the mechanics - his design gave
better performance with the same floppy media than the dumb floppy chips it
trust SSD market data?
where are we now
with SSD software?
how fast can your SSD
hidden segments in the enterprise for SSD boxes
is published by