11 Key Symmetries in SSD design - what they
are and why you need to know|
by Zsolt Kerekes,
editor - April 20, 2012
|Many of the important and sometimes
mysterious behavioral aspects of SSDs which predetermine their application
limitations and usable market roles can only be understood when you look at
how well the designer has dealt with managing the symmetries and asymmetries
which are implicit in the underlying technologies which are contained within the
Whether the designer consciously realizes that they are making a
design decision or not isn't the critical issue. The customer - who wants to
use that particular SSD in their applications environment - has to operate
within the boundaries set by those architectural symmetry limits.
symmetries are intuitively obvious - and have long been part of the filtering
process in specifying SSD shortlists.
Other symmetries are not so
obvious or commonsensical - but when you think about them in the context of -
why does this SSD work better than another? - the symmetry architecture is often
the simplest explanation behind complex operational characteristics of SSDs
which would otherwise appear to be mysterious.
There are many different
ways to design an SSD to suit the purposes of exactly the same market. And due
to the different starting points which SSD designers have in their initial IP
strengths and weaknesses - customers will see a bewildering range of design
techniques which when blended together in different combinations create usable
SSDs. There's much genuine disagreement about the best way to design SSDs and
where to put them in the apps environment.
This article isn't about
those differences. This article instead describes the key symmetries which can
be used to comparatively describe or evaluate any type of SSD using any memory
technology and any type of interface.
In an ideal world - symmetry
considerations would be on page one of the - how to design an SSD cookbook.
The fact that I've only written this article after more than 10 years
writing about SSDs and more than 20 years thinking about them - shows that the
need for a symmetry based view of SSD design has only become apparent after
reading about and mentally evaluating thousands of actual SSDs in all types of
markets and being dissatisfied with the understanding I could convey to my
readers by using other ways to describe aspects of SSD design.
reason for this late introduction is that some of the symmetry models are
actually abstractions of design concepts which didn't exist in the market or
didn't have jargon to describe them until recently. And my reason for talking
about these symmetries is to provide a practical way for readers to filter
through the chaotic range of product offerings which they will see in the real
world - rather than inventing hypothetical problems for them to worry about.
I may add more symmetries to my list later - the key symmetries in SSD design
that I will discuss in this article are as follows.
Each of these is defined and discussed in a
separate box and linked articles below.
- R/W symmetry
- power up/down symmetry
- scalability symmetry
- fault symmetry
- age symmetry
- sequential order symmetry
- application type symmetry
- roadmap symmetry
- environmental symmetry
- adaptive intelligence flow symmetry
- security symmetry
SSD design symmetry is a big
subject - whose scope will reach into every important aspect of SSD architecture
Therefore this article isn't the last word on the subject.
Instead it's my initial shot at launching a series of articles which may make
everything else I've ever written about SSDs seem just like the prologue.
|What are the symmetries in read/write
This includes such elements as throughput, latency and
related characteristics like R/W IOPS.
This is the best known of all
SSD symmetries - which is why I place it at the top of my list - to get you all
Classical enterprise storage such as
hard drives and
RAM SSDs had truly
symmetrical behavior which was a result of the underlying technologies.
contrast - nand flash (which includes most SSDs on the market today - such as
SLC, eMLC, MLC and TLC) have important R/W asymmetries which impact SSD
SSD designers work hard to disguise the strong
implicit asymmetries (which are R/W life cycles and R/W latency) by many
different design techniques.
You can read more in these classic
myths - endurance,
sugaring MLC for
the enterprise and the
write IOPS in flash SSDs.
|How long does it take for the SSD to
power down safely? |
And compare this to...
How long does it
take for the SSD to power up and be ready?
Why worry? - Because SSDs
which have bad asymmetries in this dimension prevent you from doing things which
you'd like to do.
Here are some examples
- ruin your day - an enterprise SSD which powers down correctly - but takes
40 minutes (or 6 hours) to power up and be ready - due to having to reload /
rebuild internal data.
These subjects are discussed in
my articles:- Surviving
SSD sudden power loss,
this way to the
petabyte SSD, the
Case for Fast Boot SSDs
- roadblock to new markets - if SSDs could power up as fast as they power
down - they would enable new markets.
|sequential order symmetry|
|How sensitive is the performance of the
SSD to factors like this
- the up/down direction of R/W requests into address space?
- out of order requests to the same (or contiguously proximate) adjacent
|How well does the overall SSD storage
system perform (operationally and competitively) if any one of the following
- size of the data blocks in R/W operations
- more SSDs - for example installed in the same server
- 10x more memory chips vs 10x less memory chips
Block size related performance
asymmetries are well known in the SSD world. Some SSDs are much better than
others at handling different sizes of data requests and also managing mixed size
requests sequentially. Companies which handle this well have different
marketing phrases to describe this - such as "spike-free performance"
- many more examples are possible
But it's all too easy for users to
accidentally walk into scalability symmetry traps - which relate to the number
of SSDs in the system.
They try and buy an SSD for their server. It
works great. They buy a few more - still doing OK.
A bit later down the
road they may find that they can't fill all their server or rack slots- because
of PCIe server load, or electrical power generating too much heat. So the
original SSD wasn't as scalable as they thought.
Or maybe there are
limits in the managing SSD software which mean that they can't manage it
above a certain capacity limit? Or that the speedup benefits decline after
adding more than a few more units?
A related article is:-
Why size matters
in SSD architecture
|What is the performance of the SSD with
zero faults? And how does this compare to the performance with recoverable
faults of various degrees from light (such as
or recoverable ECC)
- upto and including the failure of major subsystems.|
usually a performance penalty which creeps out of the woodwork when the accrued
level of survivable faults reaches a critical mass.
Does that mean
performance drops off to 70% of its normal level? Or can it get down
Both. These are real numbers for real
enterprise products in the market today.
Related articles are:-
and HA/ FT
Age symmetry can interact with other symmetries.
For example:- if an SSD powers down with a RAID-like fault - this may
adversely impact the power/up down symmetry - depending on how good the rebuild
/ hot-spares design is.
|How does the SSD performance change
relative to the time it has been running
You already know about degradation due to
wear-out effects and classical MTBF. Although it's good to remember that
wear-out depends on the specific type of nvm and workload and doesn't occur at
all with RAM SSDs.
- in this system?
- with this workload?
- out of the system?
But another really important age symmetry is
related to caching effects.
When an SSD (or the data set) is new -
there's no data in the caches and no knowledge about what data is hot and what
data is not. Read performance is worst. Write performance may be best. (For many
reasons such as high availability of pre-erased blocks - a
notorious factor in
early flash SSD benchmarks.)
SSD caches and
even more so in auto
tiering / SSD ASAPs the age with the data has a big impact on performance.
may be thinking microseconds or milliseconds - but actually the performance can
change (3x, 5x or 20x) when observed over time periods from tens of minutes to
hours or days - depending on how well the caching system understands the
behavior of the workload.
Finally in this category of SSD symmetries -
you may be wondering what do I mean by - Age out of the system?
means the behavior of the SSD after spending some time in an unpowered state.
Suppose you preload data onto an SSD (it may be an OS with apps for a notebook -
or a backup data set - or the control codes for a cruise missile).
you go back after a few days, a few months, 6 months or 2 years - you will see
very large differences in the data integrity on that SSD. SSDs which have high
data integrity in the always or mostly powered state may be terrible if they are
left for a long time unpowered. For flash SSDs - the cause is the difference in
remanence between SLC and various flavors of MLC. These weaknesses can be
resolved in the powered up state by healing processes - but when the power has
been off for a long time - the intrinsic defects can be significant.
- in or out of the system - is also a factor in some RAM SSDs.
out of the system - symmetries are also linked to power up / down symmetry -
discussed elsewhere in this article.
|application type symmetry|
|How sensitive is the SSD's performance
relative to different types of applications which you may throw at it?|
you've tested a new SSD for a critical project. You were happy with it. Then you
bought some more for what you considered to be a very similar role. But the SSD
performance wasn't as good as you expected.
You didn't know that but
you now suspect that something about the SSD makes it work better on some
workloads or in some racks than others. These apps asymmetries can be
- interface type - such as FC-SAN vs iSCSI
The cause is usually related to caching assumptions, write
amplification and other factors inside the SSD controller having been tweaked
and optimized for one specific type of popular benchmark to make the SSD look
good for marketing purposes - without adequate retesting to see how valid
those assumptions were for a diverse range of apps.
- apps type such as OS type, database vs web vs email vs VDI vs video on
buyers know that the indication of compatibility on a vendor's web site is no
guarantee to its actual availability or quality (if it exists). This is
especially true when it comes to how well the SSD will operate with different
Some SSD solutions only work well with a single
OS. Others work well for one OS and are mediocre or poor with others.
Now if you only ever think you're going to use one OS and one major apps type
forever you may think that this type of symmetry isn't important for you. But
think about what happens if the sole OS supplier drops or changes features in
their product which were heavily relied on by the SSD designer who only supports
that OS. An SSD with good apps symmetry - with multiple OS support - will not
be so impacted - because their design has better internal symmetries.
|You like this SSD it passes all your
tests. Then a year or so later the follow on models from the same company just
don't look competitive at all. You have to start your vendor compatibility
testing all over again.|
In SSD market history there are countless
examples of companies whose products outshone all the others at one point in
time. Then a few years later they become almost irrelevant. They are what is
known as - one trick ponies.
When choosing an SSD supplier you need to
make a judgement about how well their SSD IP and skills will hold up over
time when challenged by factors like:-
- changing memory types and process geometries
VC - you're trying to
judge how well will the company adapt to changes in the market over time.
So you have to look at their track record - or if they're a new company - ask
yourself questions like
- over reliance on roadmap success in a critical part of the design which
they don't control such as an externally sourced the embedded microprocessor
or host interface
- do they have a credible plan for a 10x faster or 10x cheaper product?
- do they have a plan which works in the memory generation 2-3 years done the
|How well does the SSD perform when the
physical environment is changed?|
Performance and operation can be
sensitive to a variety of environmental stress factors such as:-
These are standard considerations in the selection of
and military SSDs.
- RFI EMC
- ionizing radiation
You may think you don't need to worry about it in the controlled
environment of a server data center. But it can impact you there too - if you
make the wrong assumptions.
For example - in the past some vendors
attached peltier effect heat sinks to the fastest CPU chips to freeze them to
get ultimately fast performance. Should you freeze or super cool your SSD?
That depends on the memory technology. It may actually make it
slower. If there's a range of temperatures where your flash SSD runs faster or
more reliably - should you ask your vendors what it is?
|adaptive intelligence flow
|How adaptive is the SSD behavior to
changes in itself? |
Also - how bidirectional (or multi-directional) is
the ability of the SSD - when it learns one attribute of the internal SSD
internal state or data - at passing that knowledge or inference to another
part of the SSD which can use that knowledge to optimize another aspect of
performance or reliability?
All SSDs rely on processing data about the
quality of the memory as part of their normal data integrity operations. They
wouldn't work without it.
But some companies have SSD IP sets in which
knowledge about different parts of the SSD can be optimized and fed back to
control and enhance SSD functionality over and beyond the standard accepted
SSD function block boundaries.
Here are some examples:-
- PCIe SSD market -
the intelligence for managing the flash is handled by the same software stack
which talks to applications. The up/down flow of intelligence about the FTL and
the application data is really just a different view seen by the same host
processor. The ability to have data access for apps at the same latency level
as raw flash management creates many opportunities to optimize system level
behavior. ...read more
- industrial SSD
market - InnoDisk
FlexiRemap - deployed in many of InnoDisk's COTS SSDs - leverages many years of
insights into the interactions of the FTL with standard software. This has led
to a design approach within the firmware in which the responsibility for flash
management is partitioned between the CPU in the controller and the host CPU
within the driver stack. These 2 levels collaborate. The lower ones passing up
data about raw conditions and the upper levels passing down commands to trigger
certain actions inside the SSD.
|How easy is it to set up and enforce data
security in the SSD?|
How hard is it to defeat that security by using
What's the performance overhead of applying different
levels of security?
For related articles see:-
fast purge SSDs and
|How important are SSD symmetries?|
the modern era of the SSD market - users have shown an admirable willingness to
grapple with understanding many difficult concepts related to SSD design
because they know that increasing their
education and knowledge
about SSDs is their safest defense in a fast changing disruptive market
which is still experimenting with revolutionary new SSD designs which
intrinsically go beyond the limits of proven reliability and safe design rules
as part of how they advance the SSD envelope (unlike 40 years of evolutionary
processor chip and RAM
Driving this quest for deep SSD knowledge are 2 factors.
One is the tacit belief that SSDs can have magnifying
enterprise performance and cost. So users can't afford to ignore SSDs.
the other side to the SSD knowledge quest is fear of getting it
know from what they read on the web they can't blindly trust SSD suppliers
to get things right. Too many SSD vendors clearly don't understand the SSD apps
environment or SSD technology as well as they should. Many SSD vendors have
come into the SSD market because it's a fast growing
bubble with few
established market leaders
barriers to entry.
I'd like to think that in the future - the
symmetry view of SSD architecture - introduced in this article - will become one
of the ways that serious people discuss SSDs.
And as I hinted before
- "eleven" (the launch number of this key SSD symmetries article)
isn't carved in stone. I'll be adding more symmetries and more linked articles
thanks for reading this -
futures? - market analysts directory
comparing the SSD
market today to earlier tech disruptions
|"I still can't see your
- shouted Megabyte to
... from a
|I think "SSD symmetry"
is the next "endurance" - something that many SSD articles will
be talking about |
|...Editor:- discussing this article with a
leading SSD company (April 2012)|
|"Flash isn't just
about making apps run faster - in many cases it's also about being able to
guarantee that critical apps don't run slower."|
|Gary Orenstein SVP
in a webcast (May 9, 2013)|
|other suggested articles|
2.5" PCIe SSDs
SSD Companies in Q2 2012
enterprise SSD users want?
Where are we now
with SSD software?
this enterprise SSD different?
care IP including DSP for SSDs
MLC flash lives longer in my
SSD care program
SSD silos - 7 SSDs for all
enterprise datacenter needs
Efficiency = competitive
difference among leading SSDs
factors which influence
and limit flash SSD performance
meet Ken - and the enterprise SSD software
|"There is an asymmetry
in the flashs read/write/erase operations, where read and write are in the unit
of page size, usually 8KB, and erase is in the unit of block size, usually 2MB,
i.e., 256 times larger.
To reclaim a block, the FTL must copy all of the valid data in a block
to an erased block before the erase operation can be carried out, resulting in
amplified write operations.
It is well known that amplified writes,
though necessary with the asymmetry between write and erase units, can
significantly degrade the devices performance and demand a substantial portion
of the devices raw space to accommodate out-of place writes to achieve
acceptable performance, as well as reduce the life of the flash by the
SDF: Software-Defined Flash for Web-Scale Internet Storage Systems (pdf)
Editor's comments:- Baidu found that by modifying
standard SSDs to be compatible with its workload optimized Software-Defined
Flash SDF - which changes some of the management methods in the controller -
the result was 2x the usable flash capacity and 3x the I/O
recommends you read this article|
|Editor:- June 15, 2012 - I
discovered that Gareth Taube,
VP of Marketing at Kaminario
new blog in which he recommends my article about
says "Flexibility, such as being able to integrate multiple memory
technologies into a single box (like Kaminario's K2-H), is going to be
increasingly important to customers who want efficiency and customization
options. This is especially true because there are many memory innovations
coming on the near horizon." ...read
Editor's comments:- when I was writing the
symmetry article one of the things I had in mind to do was to put more examples
in it. Then I realized that having lots of examples would simply make the
One of the examples I was going to use for good
roadmap symmetry (but then forgot to put anywhere) was in fact
Kaminario - because they can leverage off whatever
Fusion-io - or later
SanDisk - does with
flash memory) and furthermore Kaminario can also leverage off whatever server
makers do with CPUs and RAM.
Roadmap symmetry is a long term consideration - important for big users who
don't like supplier churn and important for
VCs and investors too.
the kind of roadmap symmetry which I referred to above - is also the idea behind
vendors in the software defined storage market - for example
don't have to worry about the details of flash trends - because their software
and architecture leverages commodity enterprise SSDs - thereby avoiding having
to make correct long range guesses in the
enterprise SSD box