|
Predicting Key Trends in Future Flash SSD Performance
by Zsolt Kerekes, Editor
STORAGEsearch.com |
|
My thoughts re a
flash SSD
performance roadmap are summarised in this article which contains 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.
Overall
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.
The
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.
- Prediction 1 - Flash SSD throughput and IOPs (in 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. 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 read on.
|
|
|
- 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.
This needs
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 asymmetry of read to write IOPs will improve to
5 to 1 by 2010 - but not progress beyond 4 to 1. (This ratio was 10 to 1 in the
fastest devices in 2007.)
The analysis for this appears in an article
that's not yet published (Demystifying SSD IOPS). Until then - I've just
extracted some conclusions from it below.
In (typical) database
applications with Read:Write ratios of 4:1, an ideal flash SSD with 10:1 R/W
IOPS is approximately 3x (2.8x to be exact) slower in overall
applications performance than an ideal
RAM SSD with similar
MB/s throughput.
When flash SSDs improve to 5:1 R/W IOPS - the
overall applications ratio will be about 2x slower than RAM SSDs
(1.8x to be exact).
It's unlikely that the flash SSD market will try
to develop better than a 4 to 1 R/W IOPS ratio - for the reasons below.
1
-
endurance
will once again become worryingly significant for flash SSDs with this type of
architecture - even if they are SLC. And if they are
MLC - they
will burn out faster than incandescent light bulbs.
2 - if the ratio
gets much closer to 1:1 then what we are talking about is some form of RAM - or
something that's a different technology to flash anyway.
3 - the
response time (or latency) for reading small blocks of data will become a more
significant limiting factor than IOPS asymmetry - for the reasons below.
- 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
SSDs.
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
optimized).
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.
Why Z Predicts a Faster Roadmap
Acceleration than Moore's Laws
The main 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.
In the 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.
The architectural 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 developing them.
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.
MFT from
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
Link_A_Media Devices
- faster host interface controllers (and faster interfaces driven by the
needs of the SSD market rather than adapted from the HDD market)
- 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
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.
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.
Underlying
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
factors.
- Understanding
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.
- Flash
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.
- SSD
Market History - lists key market, business and technology milestones in
the 30 year history of the SSD market.
|
| . | |
|
 |
Spellerbyte's
ScryWare utility
downloaded data from his crystal ball
directly into his spreadsheet. | |
| . |
|
| . |
|
|
| . | |
