DWPD - where's it at?|
will SSDs end
you trust SSD performance data?
how fast can your SSD
control and the mythical "standard" SSD
razzle dazzling SSD cell care
and retirement plans
factors which influence
and limit flash SSD performance
the big SSD ideas to learn and unlearn in 2016?
|is it realistic to talk
about memory IOPS?|
Editor:- August 11, 2016 -
million IOPS on a single device is the title of a recent blog from Storage Switzerland which is a
briefing note about Crossbar
which operates in the conerging segments of the
alt nvm and
Among other things author - George Crump - says "Crossbar
believes that it can achieve 24 million IOPS on a single 4TB NV-DIMM without the
use of a RAM buffer or a capacitor." ...read
Editor's comments:- when startups enter new
emerging markets they are often tempted to make headline grabbing claims.
And I think the "24 million IOPS" (IOPs as in you
and I think about them) has to be interpreted in that context. (How can you
claim record breaking IOPS when all you've got is a memory IP - and that's just
part of a yet to be integrated technology set which together make IOPs.)
is not to decry the importance and validity of the
change in the SCM SSD DIMM wars market - which have consumed nearly
half of my working hours in the past year.
We saw similar wild
claims when the startup Fusion-io
was trying to get across how
PCIe SSDs would change
the enterprise storage market by reference to the nearest similar technology
when Fusion said in 2007 it would replace SANs. (Because SAN based SSD
accelerators were at that time the SSD market's dueling weapons of choice.)
back to Crossbar - there is a genuine problem for the industry (which I touched
on in an earlier post about Diablo's
DMX software) - which is - what are the most useful metrics to judge tiered
memory systems by?
As we've seen in the
SSD accelerated storage
pool market since 2009 - there's a wide spectrum of use cases and cost
considerations which have many viable business intersections.
need new "goodness" numbers for DIMM wars memories.
think using IOPS to characterize a memory product is less useful to describe why
people might want to look at it than wattage, raw capacity in a DIMM,
uncached raw R/W latency and price.
And - most important of all -
what software does it work with? And how well does the software behave?
Non-Balanced Wear Leveling - a paper by Renice
July 28, 2016 - Renice
Technology recently published a paper -
Wear Leveling Algorithm (pdf) - which outlines the thinking behind a
specific technique in its industrial SATA3 SSD controller - model
- to improve
upto 3x compared to traditional methods.
This is one of the
several techniques used in this controller to overall get a 20x
improvement in lifespan when using MLC. ...read
the article (pdf)
Editor's comments:- Ever since the
first flash devices were evaluated it has been known that some blocks are much
better than others.
As an example in this paper Renice shows that
in a modern 16GB MLC flash chip - even after just 10 P/E cycles the controller
is able to see a 3x difference between the fastest and average program time and
over 30% difference between the slowest and fastest read times.
quality of wear resistance tells you something which can be used to grade
Renice's non-balanced wear leveling algorithm leverages
these naturally occurring process variations so that "the higher
wear resistance blocks are selected to be erased more times while the lower
ones get protected instead."
Although there are no fundamentally
new ideas presented in this paper - because the technique is just one
permutation of many from the superset of all
techniques - this paper does provide a useful survey of classical wear leveling
techniques along with their associated trade offs in performance and endurance.
I got a good sense of judgment and balance in this paper - given the
Context is always important - and these techniques
are discussed in the context of general purpose, simple low power
which use modest speed
That's distinct thinking from new generation enterprise
array controllers in which visibility into other SSDs in the same array,
larger ratios of DRAM and knowledge about the applications software stack can
also be leveraged to reduce endurance.
an SSD way of looking at hard drives
Editor:- May 4,
2016 - In an ironic twist of fate - it looks as though hard drive vendors may
find it useful to characterize some aspects of HDDs in a way which can be easily
related to value judgement numbers created for SSDs.
A recent article -
did HDDs get SSD-style DWPD ratings? - in The Register brings to our
attention that hard drrives are now being specified with write limits and the
author Chris Evans
(who also blogs as Architecting IT )
conveniently provides readers with a list of HDD models along with their DWPD
equivalent ratings. ...read
what's the state of DWPD in
worst case response times in DRAM arrays
March 1, 2016 - Do you know what the worst-case real-time response of your
electonic system is?
One of the interesting trends in the computer
market in the past 20 years is that although general purpose enterprise servers
have got better in terms of throughput - most of them are now worse when it
comes to latency.
It's easy to blame the processor designers and the
storage systems and those well known problems helped the
SSD accelerator market
grow to the level where things like
PCIe SSDs and
hybrid DIMMs have
become part of the standard server toolset. But what about the memory?
memory based on DRAM isn't as good as it used to be. The details are documented
in a set of papers in my blog -
loving reasons for fading out DRAM in the virtual memory slider mix.
1mS VCC blip to 0V is enough to kill many SSDs!
January 27, 2016 - an incoming email this morning started like this...
Following on from our previous conversation on power
problems - we've found that a significant number of SSDs will die entirely if
the voltage rails are pulled to ground for (as little as) a 1mS period.
sales demos often go along the line of - yes you can do that, the drive should
handle it fine....... oh crap :-)
The sender was Andy Norrie, Technical
Director - Quarch
Technology who was continuing a conversation started 3 years earlier
about problems in SSDs due to untested power loss vulnerabilities.
company designs intelligent power
rail units for SSDs which can help designers verify imunity from, or
sensitivity to, what-if? power rail disturbance vulnerabilities which are
caused by scenarios like hot plug spikes, power up-down sensitivity, noisy
today that it is becoming better known by SSD reviewers and Andy also told me
- "last month we put around 75 test units into a single lab of a major SSD
...Later:- More detail on the exact nature of
the 1mS blip test emerged from Michael Dearman,
Founder of Quarch Tech who (after seeing this post) said this...
All, I have broken several drives in customer demos with the 1mS test, the key
factor is that in this test we don't just stop supplying power and let the rail
float (like a disconnection), we pull the rail hard down to ground (as would
happen with a power supply crowbar). In this instance the drive should isolate
itself from the host power supply to preserve its internal charge and complete
its power down, but doesn't always manage it!"
On seeing that - Sudhir Brahma Principal
Engineer at Dell
said:- "Usually these glitches result in data corruption....I will be
surprised if they kill the ssd. I was working on one such issue- traced it down
to a firmware area (a bug) where a glitch could potentially cause loss of data
in flight......u need to write intrusive code for that....true it is tough Are
u saying that by doing those glitches, u actually stoned the drive? We built
such glitch generators and ran them on the ssds, but never had a stoned SSD."
Dearman said - "yes, we have had production drives go completely dead."
were other interesting comments re the above post from people who have also
used Quarch Tech's SSD testers - which you can see via my
MIT research findings -FPGA processing in flash arrays
July 14, 2015 - Flash SSDs with in-situ processing in
cached servers can deliver nearly the same apps performance as fat RAM
servers (but at much lower cost and lower electrical power).
one inference from a recent story -
cost and power consumption for big data - in MIT news - which summarized a
research paper at ISCA
2015 - BlueDBM: An
Appliance for Big Data Analytics
Part of the system architecture in the research included a
network of FPGAs which routed data to the flash arrays and offloaded some of
the application specific processing.
This is not a replacement for
DRAM said Professor Arvind whose group at
MIT performed the new work. But there may be
many applications that can take advantage of this new style of architecture.
Which companies recognize: Everybodys experimenting with different aspects of
flash. Were just trying to establish another point in the design space. ...read
less than 10% of FC SAN sites rely on 3rd party benchmarks
June 9, 2015 - Load DynamiX
(a storage performance
testing and validation
company) recently released the
of a survey (pdf) characterized by heavy users of FC SANs (71%) and 2PB or
more of data (76%).
Among the findings in this set of 115
- over half (54%) planned to add all flash arrays to their storage assets in
the next year
- one third (34%) used custom performance scripts as part of their pre
purchase and deployment evaluations
the article (pdf)
- users were heavily reliant on their current and potential vendors for news
about new products and technologies - and nearly twice as likely (83%) to
rely on news from vendors compared to online magazines (44%)
Benchmarking and Performance Resources
February 6, 2015 - When it comes to SSDs - an SSD which is
faster in a way
that you can economically use - such as by converting faster
competitive dollars (trading banks) or by satisfying more virtual users with
less servers (nearly everyone who owns a lot of heavily used servers) is
worth looking at.
not the only thing (and often is not even the most important thing) which makes
up the cost of buying
an SSD - or the
to buy it - performance has been one of those parameters which - because it
has helped to sell products - even when the
numbers were unreliable
or abused - has attracted a great deal of creative literary output in the
SSD industry. Most of it fiction. Some of it fact.
I've written a lot
of articles and emails on this theme myself. So many indeed - that I sometimes
find myself in danger of writing something new - and then getting a sense of
deja vu. IOPS?
- I've got a feeling I wrote something like this before? A quick search confirms
- yup I did. - Was it yeally that long ago? Let's just update the links so it
makes sense if someone else finds it later.
It seems I am not alone
in that respect. And a recent post on linkedin suggests a much better way of
The idea came from Greg Schulz, Founder of
StorageIO - who has recently
curated a whole bunch of articles which he's written, edited or likes into a
single resource page - which he calls - Server
and Storage I/O Benchmarking and Performance Resources
have the time - Greg has many articles on this topic which will inform and
hot topics at StorageSearch.com?
12, 2015 - StorageSearch.com
reader interests in Q4 2014 are reflected in the
most popular SSD
The big change - compared to a year ago - is that
SSDs have become as hot a subject for reader research as
PCIe SSDs were back
in an earlier phase of the market in the first half of
Who Needs 10Gbps USB?
Editor:- October 29, 2014 -
SSDs are at the forefront of the thinking in a new article today -
Really Needs USB 3.1? by Eric
Eric's blog sketches out a 5 years into the future application
picture for this new (10Gbps) iteration of the USB connected story.
also:- storage interface
chips, market research
Tanisys enters SSD ATE market
Editor:- August 5, 2014
- Tanisys Technology
details of a new SSD ATE test system which will be shown at the Flash Memory Summit.
Tanisys's SX3-OGT test
system (which includes benchmarking and validation suites from
supports popular SSD interfaces including
SX3-OGT also supports fast emerging protocols such as NVMe and AHCI. The SX3-OGT
is available in bench top configuration for engineering applications and with
multiple burn-in chambers for production.
And the best buy SSDs shall be the worst (if you change your
Editor:- August 2, 2014 - An applications optimized
SSD system can be the cheapest buy - if you always use it for the original
purpose - but it can be a poor choice if you throw the wrong type of
applications at it. Enter - the good ole general purpose fast SSD array.
conflicts are examined in a new blog -
Flash Storage Systems multi-task! written by Woody Hutsell,
IBM who among
other things says - "It just so happens that flash appliances with
built-in deduplication are the worst choices for database acceleration."
The idea that an SSD which is best for one type
of use may have the worst characteristics for another - was also examined from
an architectural point of view in my classic article -
how fast can your
SSD run backwards?
real-world performance of flash storage systems
July 23, 2014 - Editor:- July 23, 2014 - How does flash storage perform in the
real world? - Demartek
aims to provide some answers by reporting on the performance tests
which it has carried out on SSD and hybrid systems from many of the leading
enterprise SSD companies in a session next month at the Flash Memory Summit (August
Demartek says attendees will come away with reasonable estimates
of what they can expect in practice and the results also reveal additional
advantages of flash-based storage, with what Dennis Martin, President
- Demartek calls "happy side effects".
(later) comments:- see Dennis's paper -
Performance of Flash-Based Storage Systems (pdf)
say hello to CacheIO
Editor:- June 10, 2014 - CacheIO today
announced results of a
benchmark which is
described by their collaborator Orange
Silicon Valley (a telco) as - "One of the top tpm benchmark results
accelerating low cost iSCSI SATA storage."
CacheIO says that the 2
million tpm benchmark on CacheIO accelerated commodity servers and storage
shows that users can deploy its flash cache to accelerate their database
performance without replacing or disrupting their existing servers and storage.
Editor's comments:- The only reason I mention this otherwise
me-too sounding benchmark is because although I've known about CacheIO and
what they've been doing with various organizations in the broadcast and telco
markets for over a year - I didn't list them on StorageSearch.com before.
was partly because they didn't want me to name the customers they were working
with at that time - but also because with
SSD caching companies
becoming almost as numerous as tv stations on a satellite dish - I wanted to
wait and see if they would be worth a repeat viewing. (And now I think they
what's in a number?
Editor:- March 4, 2014 -
is a latency based configuration metric - proposed as a new standard by StorageSearch.com - which can
tersely classify any enterprise server - as seen from an SSD software
perspective - by a single lean number rating from 0 to 7. ...read the article
ESG reports on clustering Virident's PCIe SSDs
January 23, 2014 - ESG
today published a
report for Virident
/ HGST's FlashMAX
II (PCIe SSDs)
which validated the ability of this product family and its
(vShare ) to be configured for useful operation in an Oracle RAC environment
availability configurations which were clustered via
ESG said performance was good they also commented on some current
limitations of the product suite for this type of application. In particular:-
- the lack of a graphical interface for setup and performance mintoring,, and
- the lack of support for other supported fabrics such as 10GbE (mentioned
in the report as a future option), and PCIe fabric (which was not mentioned at
all in this report).
how to avoid hot pluggable PCIe SSD failures
December 3, 2013 - What happens if you test
PCIe SSDs for their
sensitivity to data
corruption or even failure
- in the event of
You'd think that with more hot pluggable products coming
into the market -especially in the
2.5" form factor
- that the experimental outcomes would be known by the designers and problems
debugged so that users wouldn't have to worry.
In August 2013 Quarch Technology
launched some special
test equipemnt to inject power related faults into PCIe SSDs - and the
company today announced
it has extended this range to automate power line error testing of PCIe SFF
Technical Director at Quarch Technology told me today that "Almost every
combination of test kit we have tried in Quarch (a number of friendly customers
lent us kit and eval drives to get the new module up and running) has failed in
some way. Sometimes failing to come back up again nicely, sometimes with a full
BSOD which will almost
certainly have risked data loss."
memory channel SSD vs PCIe SSD write latency - 3rd party
Editor:- November 7, 2013 - If you've been wondering
storage compares with PCIe
here to a new whitepaper (pdf) which includes some useful data. The
application isn't important but it's the first public glimpse which goes
usefully beyond the graphs shown in the product launch documents.
highlight for me is - a mean write latency of about 30µS for MCS compared
to about 100µS for PCIe SSD- at a particular R/W ratio which may of
course look nothing like your own setups.
Nimbus scores $40 per virtual desktop in IOmark-VDI
October 1, 2013 - I was talking recently to Tom Isakovich,
CEO and founder of Nimbus
Data Systems about the state of the market and a possible date for an
IPO (which you can read more about in
that we moved onto the subject of
customers buy SSD arrays - and we traded stories about some of the
explanations which get tossed around like
IOPS per dollar -
which when you scrutinize them in any detail are ridiculous. We've both seen
leading edge silicon SSD companies put nonsensical graphs into their marketing
presentations which don't lead you anywhere useful in the real world if you
follow the superficial analysis. (That's because these vendors don't make
systems - and are many steps removed from genuine enterprise thinking.)
said most of his customers couldn't tell you how many
apps were demanding.
I said I've been writing about the "cost
of satisfying a given number of virtual users for a particular type of app"
as being a useful comparison figure (for storage). We both agreed that even if
enterprises don't know for sure what their throughputs or IOPS are - they have a
good idea of how many users they're trying to serve within their organization or
at customer facing web interfaces. The payroll tells you one, the marketing
people can tell you the other. And accounting can tell you how much it all
costs. You don't need storage
analyzer tools to get a feeling for where the ground level lies.
our conversation had already veered towards the subject of a simple way that
users can compare the costs of SSD storage for particular types of apps - and
as I'd asked the question - he said there was a benchmark called IOmark-VDI
which Nimbus had participated in recently with the
He said he went into the process because he thought it might be a good thing to
try out - and was gratified to found out that it shows the Nimbus product in a
very good light with a
under $40 per virtual desktop (pdf) achieved by a 2U Nimbus system
supporting up to 4,032 linked clone VDI images.
measuring enterprise SSD performance - intro by Micron
August 23, 2013 - EDN today
published a introductory article on the subject of
enterprise SSD performance - written by Doug Rollins, Senior
Applications Engineer at Micron
- which could be useful for newcomers to this topic as it expounds some of the
basic assumptions and jargon. ...read
Can you trust flash SSD
specs & benchmarks?
FIO's ION software enables Breakthrough Shared Storage
Editor:- June 13, 2013 - The performance of
Accelerator software - which you can add to its PCIe SSD cards, any
standard server and some FC adapters to roll your own SAN
rackmount SSD -
is the point of a new
blog by the company
today which celebrates recent benchmarks for
2, 4 and 8
processor HP server configuartions (pdf).
Stec's profiler removes guesswork in sizing SSD caches for
Editor:- May 21, 2013 - Stec today
that it's offering a free profiling tool -
Profiler - which can enable users to determine how much benefit they would
get from using its
(SSD caching software) - before they even install any SSDs.
company says that the "non-disruptive installation" can save hours of
administrative trial and error by recommending the optimal block size, and the
capacity and type of SSDs to be used for maximum performance gain.
SSD performance characteristics and limitations
March 15, 2013 - published today - the new home page blog on StorageSearch.com is -
a toolkit for
understanding flash SSD performance characteristics and limitations.
brings together in one place many of the tools I use every day when thinking
about and assessing SSDs.
in memory database even better with FIO's flash
November 20, 2012 - McObject
recently released new
benchmark results which indicate that the in-memory database company is not
so unfriendly to flash SSDs as you may have thought from reading earlier
company positioning papers.
It seems that a software product - which
was originally designed for the DRAM-HDD world - is a good fit in the flash SSD
world too - if you have the right scale of data and the right SSD. ...read more
new article - adaptive flash care IP (including DSP)
June 19, 2012 - A few months ago I promised readers that I would publish a
tentative list of SSD companies who use what I loosely called
DSP technologies in SSD IP" in their new designs. .
of the most important design techniques being used in some leading flash SSDs
- in which the SSD designer can adapt the reliability, speed and power
consumption of the SSD - not based on some faw away population model of flash
chips - but optimized for the chips in each SSD - and adapting the
controller behavior to
what is measured and learned from interacting with the flash chips installed.
This is a market changing technique.
suite could speed up auto-tiering SSD evaluations |
|Editor:- November 29, 2011 -
availability of its Disk I/O Ranger software analysis tool for Windows
The company says this will help users diagnose and
understand disk storage access performance problems and to to verify that QoS
levels are being met at the application/file/device level. It could also
simplify the evaluation of
appliances by collecting real-time metrics.
I asked Tom West, President of hyperI/O what he was seeing of the SSD
market from his perspective of selling storage analysis tools. He said -"One
of the major users of the hIOmon software is listed within the top 10 of your
latest - Top 20 SSD
Microsemi reports shake rattle and roll SSD results
May 19, 2011 - Microsemi
today announced that its TRRUST-STOR
(2.5" rugged SSDs) are the industry's first SSDs to pass zero-failure
testing at vibration
levels that are consistent with the industry's most severe environments.
"No other SSD manufacturers have published zero-failure results
at this level of vibration testing, which was conducted while our drives were
fully operational, reading and writing data," said Jack Bogdanski, director
of marketing for Microsemi. "The ability for SSDs to perform flawlessly
under adverse environmental conditions is becoming increasingly important for
applications where it is critical that data be protected at all times."
Microsemi's SSD units were pre-conditioned at 85°C for 336
How and why to monitor VM Performance
February 23, 2011 - How
to Proactively Monitor VM Performance is a new article on
Data Center POST written
by Alex Rosemblat,
Product Marketing Manager at VKernel
- who says "Proactive monitoring of a virtualized data center can assist
in finding potential performance problems before they occur..."
comments:- OK he says a lot more than that - and that's why I mentioned his
I used to do a lot of performance analysis in my
pre cut and paste
career because I designed systems with guaranteed apps response times. And
in my current job I always check my stats before I look at my email. So I have
a lot of empathy for the storage
test and analysis market. The more you understand about the internals of
complex systems the less likely you are to get mugged by them. ...
read the article
SandForce publishes list of approved test tool partners
January 31, 2011 -SandForce
has started a directory of companies, tools, technologies and services to
help SSD designers integrate its
SSD processors and get
them to market more quickly.
Each member company in the new
Trusted program ensures that their products and/or services fully
support SandForce SSD Processors and provides response to SandForce customer
inquiries within 24 hours while committing to high-priority support for fastest
Editor's comments:- 6 out of the 7 initial
companies in the new program provide
test / design verification
||Megabyte liked to keep his systems
in peak condition.|
|The more you study the
characteristics of different SSDs - the quicker and more easily you will start
to anticipate useful behavioral characteristics of any new SSD - and assimilate
new SSDs in your plans. And you'll start to recognize symptoms of "missing
technical information" too. These are things which it's important for you
to know - but which don't appear in the initial info you see about the new SSD.
|understanding flash SSD
performance characteristics and limitations - a toolkit|
| "Our past work
showed that application-unaware design of memory controllers, and in particular
memory scheduling algorithms, leads to uncontrolled interference of applications
in the memory system" - said Onur Mutlu, Carnegie Mellon University.|
|Are you ready to
|The paradox for the
potential use - is that 2 different flash SSD designs which have apparently
identical throughput, latency and random IOPS specs in their datasheets - can
behave in a way that is orders of magnitude different in real applications. |
|the Problem with
Write IOPS in flash SSDs|
|in-situ SSD processing
This is about closing important gaps in the intelligence of message
passing and the speed of data access between application processors and SSD
|10 key SSD
ideas which became clearer in 2014|
management in flash SSDs|
|This is an introduction to the thinking behind
one of the many vital functions inside a flash
media defect quality in new flash memory chips has grown steadily worse in the
past 10 years as geometries have shrunk.
||This article enumerates
the scale of the problem and explains how intrinsically dodgy flash memory
is transformed into dependable flash SSDs which you can entrust with your data.