Predicting and Reporting Key Trends in Flash SSD Performance - upto
|I'm often asked
where I think the
flash SSD market
is heading in terms of performance. |
Although my earlier article
RAM SSDs versus
Flash SSDs - which is Best? looked at the state of the art (in
didn't include a performance roadmap.
This article does. And it has a
bunch of simple to remember predictions or "laws". For convenience
you can think of them as "Zsolt's Laws". Or maybe "Z's Laws"
is easier to spell. The first law is the only one you need to remember.
this article tells you the market model assumptions I would work to if I were
setting up a new
SSD company to design the
fastest flash SSDs.
Obviously I'm not doing that - because otherwise the many people I
talk to in SSD companies around the world wouldn't be talking to me.
I've got a good crystal ball and the lucky thing about my many past
predictions related to the storage
market - is that if they weren't going to be true at first - then suggesting
to companies that they ought to spend more time looking at things in a
particular kind of way becomes a self fulfilling prophecy. My safety net is that
as many designers in the industry also read these articles - if I say something
that's too outrageously wrong - they correct me - before I look too ridiculous.
A quick delete, cut and paste, and it's even more accurate than it was before.
rate of speedup predicted is much faster than would be predicted by Moore's Laws
- and that's the significant point. Some explanations why - are included at the
foot of the article.
The single most useful thing to take away from this
article is the assertion above. But if you want to have some more to think about
- Prediction 1 - Flash SSD throughput and
traditional HDD form factors) will more than double every year in the period
from 2007 to 2012.
This predicts (in effect) that in 2011 a single
3.5" form factor flash SSD will be able to deliver similar throughput to
some of the fastest RAM SSDs available in 2007, with over 2,000MB/s sustainable
reads and writes.
- Prediction 2 - Rackmount flash SSD throughput and IOPS performance
will be a multiple of the performance for a single disk. These factors have
already been shown to be scalable in SSD RAID arrays.
little explanation as some of the results are intuitive and we've already
published plenty of articles on this subject. However, some of the
architectural features which are now used in SSD RAID systems - such as
MFT technology - can also be designed
into individual SSD disk modules.
- Prediction 3 - The
sustained read to write IOPs will improve from 10 to 1 (the fastest devices
available in 2007) - but will never achieve parity (1 to 1).
change occurs in the market flash SSD arrays will become viable choices in many
enterprise server speedup applications which hitherto had been the exclusive
domain of RAM SSDs.
In (typical) database applications with Read:Write
ratios of 4:1, an ideal flash SSD with 10:1 R/W IOPS is
approximately 3x slower in overall applications performance than an
RAM SSD with similar
...Later:- I was wrong about this one.
Violin Memory designed
a flash SSD with balanced random R/W IOPS in Q4 2008 using a
combination of factors.
and fast garbage collection ensured a constant flow of pre-erased flash.
- their non blocking RAID
RAM cache architecture enables a read to be done from a virtual flash block
while an erase-write is still in operation on the same flash block.
in April 2009 - SandForce
Fiction - all flash SSDs have unbalanced R/W performance
. This is one of a set of articles which discuss design factors inside
their small form factor SSD controller.
- Prediction 4 - Latency in flash SSDs will not scale in the same
way as throughput, and will always be significantly worse than that in ideal RAM
The ratio of read access times for RAM SSDs compared to flash
SSDs may improve for a few years (as the gap gets smaller) but then it will
hit a brick wall - and may in fact get worse again.
The reason is -
that flash SSDs have not yet been optimized for latency - so there is some scope
to reduce the latency gap with RAM systems (which have already been highly
But in future product generations as flash SSDs
increase in density - a read or write cycle becomes an increasingly complicated
on-chip process - which includes calibration, error correction and address
translation all being done by controllers between the memory array and the host
interface controller or card data bus.
This series of steps (to do a
simple read) will diverge from what happens in a typical RAM to the point
where flash and RAM look like completely different species. That's unlike
earlier generations of flash in which the read cycle looked the same as a
static RAM - but simply took longer.
Here are some other preliminary notes of explanation.
Predicts a Faster Roadmap Acceleration than Moore's Laws
reason that the flash SSD market will deliver faster products much sooner than
predicted by Moore's Law is that 2 additive factors are at work in this phase of
the market:- architecture and semiconductor process technology.
past there was little point in manufacturers integrating very fast architectural
features into flash SSDs - because they added to the cost - and there wasn't a
big enough established market to buy them.
technologies that can speed up performance and IOPS were originally
independently developed by various oems to suit particular products or markets.
Until the SSD market reached a critical mass where enough users signalled they
would buy faster products if they were available - there was no point in
the key architectural features which will increase
throughput and shrink the asymmetry gap in read / write IOPS are:-
- parallelization of the internal media arrays
- improved media management technology.
EasyCo is a software
solution which has already been mentioned above. But the same algorithms could
be run in the SSD hardware.
Another licensable solution - which is
already being developed at the chip level is IOP Buster architecture from
- faster host interface controllers (and faster interfaces driven by the
needs of the SSD market rather than adapted from the HDD market)
Scaling any one of the factors above
requires significant investment in IP. There is also significant risk that the
overall balance of the product specs which results doesn't match the market's
expectations for price and performance at the time.
- hybridizing on board memory technologies - for example using faster
RAM-like non volatile memory in some parts of the device and slower flash-like
memory in the bulk storage arrays
A lot of trial and
error will be involved as oems throw products at the market which tweak the
technologies they understand best - and see which products stick.
all the architectural improvements - there will also be evolutionary and
revolutionary changes in semiconductor processes occurring at the same time.
Some of these will enhance currently known architectures, while others may make
some architectural features obsolete.
By around 2013 - the flash SSD
tornado should have reached a point where the architecture of an ideal SSD is
well established - and the ongoing developments will be drive more by process
changes than anything else.
We're going to publish hundreds of articles
about this subject as the roadmap takes shape. So don't expect to see all the
answers in a simple note like this. But I'll add more notes and links to this
article as time goes on.
Here are some related articles.
- the Fastest
SSDs - updated regulalry - this gives the state of the art in popular form
Flash SSD Performance (pdf) - by Douglas Dumitru, CTO EasyCo LLC is a
reference for systems engineers who want to understand the dynamics of flash
versus hard disk performance running real applications.
in the Enterprise (pdf) by Jamon Bowen, Texas Memory Systems describes
some the properties of Flash memory and then explains how their RamSan-500
product leverages its strengths and compensates for its weaknesses to offer the
fastest enterprise ready rackmount Flash SSD system.
Tradeoffs for (SLC flash) SSD Performance (pdf) by Nitin Agrawal, Vijayan
Prabhakaran, Ted Wobber, John D. Davis, Mark Manasse and Rina Panigrahy looks
in detail at the internal architecture of a typical commercially available SLC
flash SSD and the various points where choice of software algorithms can impact
- Why I Tire of -
"Tier Zero Storage" - You don't need to waste any of your
precious brain cells by investing "tier 0 storage" with an importance
this travesty of storage jargon really doesn't deserve.
Market History - lists key market, business and technology milestones in
the 30 year history of the SSD market.
|the 3 fastest flash
|Are you tied up in
knots trying to shortlist flash SSD accelerators ranked according to
published comparative benchmarks?|
You know the sort of thing I mean -
where a magazine compares 10 SSDs or a blogger compares 2 SSDs against each
other. It would be nice to have a shortlist so that you don't have to waste too
much of your own valuable time testing unsuitable candidates wouldn't it?
StorageSearch's long running
fastest SSDs directory
typically indicates 1 main product in each form factor category but those
examples may not be compatible with your own ecosystem.
If so a
new article -
the 3 fastest PCIe
SSDs list (or is it really lists?) may help you cut that Gordian
knot. Hmm... you may be thinking that StorageSearch's editor never gives easy
answers to SSD questions if more complicated ones are available.
||But in this case you'd be
wrong. (I didn't say you'd like the answers, though.) ...read the article|
sudden power loss|
|Why should you care
what happens in an SSD when the power goes down? |
This important design
feature - which barely rates a mention in most SSD datasheets and press releases
- has a strong impact on
SSD data integrity
This article will help you understand why some
SSDs which (work perfectly well in one type of application) might fail in
others... even when the changes in the operational environment appear to be
|RAM Cache Ratios in
|Editor:- July 28, 2009 - StorageSearch.com today published
an article - which proposes new terms to describe and differentiate
products in the flash SSD market.|
|| It is hoped that the new classification
jargon will be useful to users who have to evaluate lots of products, and will
be useful to vendors as a shorthand when communicating about
different segments within their flash SSD product lines. ...read the article|
Unveils New flash SSD Controller |
|Editor:- April 13, 2009 - SandForce today
from stealth mode and unveiled its
SF-1000 family of SSD
Processors - aimed at oems building SATA flash SSDs.|
Its 2.5" SSD
reference design kit is the fastest 2.5" SATA flash SSD on the market -
with 250MB/s symmetric R/W throughput and 30,000 R/W IOPS.
are expected to release both SLC and MLC flash-based SSDs using SandForce
single-chip SSD Processors later this year
"With a deep
understanding of both system- and silicon-level issues, we've integrated the
right balance of reliability, performance, power, cost, and time-to-market in
our SSD Processors while supporting multiple flash vendors' technology,"
said Alex Naqvi, President and CEO of SandForce. "Our products combine key
processing elements with hardware automation to efficiently address the
traditional shortcomings of flash memory. This allows OEMs to provide
enterprise-class SSDs to the mass-market using both SLC and lower-cost MLC flash
devices while delivering peak read and write performance throughout the drive's
Editor's comments:- I asked SandForce's President & CEO,
Alex Naqvi, for more details about the various package of technologies which
are bundled in the company's "DuraClass Technology" - which achieves
impressively high IOPS without relying on over-provisioning or large external
RAM caches. In particular I wondered what part, if any its choice of processor
SoC (from Tensilica) had to play.
Naqvi explained - DuraClass performance doesn't come from the choice of
processor - but in the way that they have integrated various design techniques
with very fast hardware (proprietary chips) which the company has designed to
accelerate the core bottleneck functions of a flash SSD controller.
||In concert with other techniques, such as the
ability to reorder data before it is written to flash (thereby attenuating
write endurance by 2 orders of magnitude), RAID like internal protection and
very fast garbage collection SandForce's DuraClass Technology results in small
form factor enterprise class flash SSDs which have no daily write limits for
MLC flash and symmetric R/W IOPS. |
|the Problem with
Write IOPS in flash SSDs |
|the "play it again Sam"
Flash SSD "random write IOPS" are now similar
to "read IOPS" in many of the
why are they such a poor predictor of application performance?
why are users still buying
RAM SSDs which cost an
order of magnitude more than SLC? (let alone
MLC) - even
when the IOPS specs look similar.
||This article tells you
why the specs got faster - but the applications didn't. And why competing SSDs
with apparently identical benchmark results can perform completely
|Can You Trust Your Flash
|Editor:- I've noticed is that the published specs
flash SSDs change
a lot -from the time a product they are first announced, then when they're
being sampled, and later again when they are in volume production. |
the headline numbers get better, sometimes they get worse. There are many good
reasons for this.
The product which you carefully qualified may
not be identical to the one that's going into your production line for a
variety of reasons...
|| And here's another thing to worry about...
The enterprise flash SSDs which you benchmarked yourself - may
surprise you by running much slower when deployed in your own
applications due to common "halo" errors which are implicit in the
set ups of many performance test suites which were originally designed for
HDDs. ...read the