From: David Brown on
Arno wrote:
> David Brown <david.brown(a)hesbynett.removethisbit.no> wrote:
>> Barry OGrady wrote:
>>> On Sat, 6 Feb 2010 22:13:20 -0800, "trs80" <trs80(a)YAHOO.COM> wrote:
>>>
>>>> Does anyone yet make a TB Flash memory in a 3.5" drive physical format. If
>>>> so, could you pass on a reference? The interface would need to support
>>>> about at a 400MB/s sustained rate. I can work with any interface such as
>>>> Fiber Channel or whatever .
>>>> thanks for any tips
>>> Its not available yet, and remember flash ram is slow and has limited
>>> read write cycles.
>>>
>
>> Flash drives are not normally considered "slow" - read performance of
>> 200+ MB/s, and writes of maybe 70 MB/s are possible with good drives.
>> That's a lot faster for reading than even top-range hard disks, and
>> similar for writing (the OP doesn't specify if they want reading or
>> writing speeds).
>
> Indeed. however a RAID0 with relatively cheap disks will give you
> 200+MB/s read and write speeds for large accesses. A 4-way
> RAID0 (possible at least with Linux software RAID and likely with
> the xBSDs as well) should reach 400MB/s large access speed.
>

I agree that raid is the cost-efficient way to get this sort of
bandwidth (the cost-inefficient way being a PCI slot ram disk).

> For small accesses the story is different, onlt a RAM frontend will
> reach this speed here, FLASH may be even slower than disk here,
> especially on write.
>

Small read accesses are extremely fast with flash disks - far faster
than with hard disks, since small reads are dominated by the seek times.

For small writes, it is certainly true that these are slower to complete
on a flash disk than on a hard disk, and less efficient than streamed
writes. But outside of synthetic benchmarks, so what? Small writes are
cached by the OS - as far as the application is concerned, they happen
almost instantaneously. And as long as you don't have too many of them
flushing to the disk at the same time, the writes are not going to cause
other performance issues for the flash disk - if the disk needs to read
from the same flash chip, the erase/write can be paused temporarily.
You'll only see a real-world problem if you are dealing with an
application that makes small writes and lots of fsyncs, combined with an
older flash drive that is poor at hiding the garbage collection.

>> I agree about the sizes - you don't get flash drives as big as 1 TB at
>> the moment (at least, not in standard 3.5" formats).
>
>
>> As for write endurance, you are about a decade out of touch...
>
>> <http://www.storagesearch.com/ssdmyths-endurance.html>
>
> Well, not for USB flash. I did recently torture a 2GB Kingston
> USB to death and it had consistent data errors (with no error
> meassege to make matter worse!) after about 3500 full overwrites.
>

USB flash devices are generally optimised for low costs rather than high
quality or high endurance. They also often have very poor erase block
management, since they have few chips and also must minimise the risk of
data loss if the device is removed unexpectedly. This means a single
block write to the device can cause many erase/writes to the flash.

> Say the OP wants to overwrite his disk at 400MB/s, then 3500
> full operwrites are reached after about 100 days of operation.
>

He doesn't say whether he wants to stream reads or writes, or how long
he wants to sustain the transfers. My guess would be that he'd like to
break off writing and do the occasional read - there are not many
applications which produce new data at a rate of 34 TB per day, nothing
of which needs to be kept for more than forty minutes!

> I expect SATA FLASH is better, but not all may be. What however
> does not happen with modern FLASH is that writing a few 1000
> times to a single location kills the drive. Traditional FLASH
> without wear leveling had that problem, with some dying
> after 10000...100000 writes to the same sector.
>

Modern SLC flash chips will have endurance in the range of at least a
million erase/writes if you are nice to them (i.e., keep them at room
temperature). MLC devices used in cheaper disks have significantly
lower endurance.

So for continuous writing using good SLC disks, he's got an average of
something like 80 years before write endurance is a problem.

From: trs80 on
great inputs. thanks for the help
"trs80" <trs80(a)YAHOO.COM> wrote in message
news:4Ssbn.22419$aU4.3835(a)newsfe13.iad...
> Does anyone yet make a TB Flash memory in a 3.5" drive physical format.
> If so, could you pass on a reference? The interface would need to support
> about at a 400MB/s sustained rate. I can work with any interface such as
> Fiber Channel or whatever .
> thanks for any tips
>


From: Arno on
David Brown <david.brown(a)hesbynett.removethisbit.no> wrote:
> Arno wrote:
[...]
> Small read accesses are extremely fast with flash disks - far faster
> than with hard disks, since small reads are dominated by the seek times.

True.

> For small writes, it is certainly true that these are slower to complete
> on a flash disk than on a hard disk, and less efficient than streamed
> writes. But outside of synthetic benchmarks, so what? Small writes are
> cached by the OS - as far as the application is concerned, they happen
> almost instantaneously. And as long as you don't have too many of them
> flushing to the disk at the same time, the writes are not going to cause
> other performance issues for the flash disk - if the disk needs to read
> from the same flash chip, the erase/write can be paused temporarily.
> You'll only see a real-world problem if you are dealing with an
> application that makes small writes and lots of fsyncs, combined with an
> older flash drive that is poor at hiding the garbage collection.

>>> I agree about the sizes - you don't get flash drives as big as 1 TB at
>>> the moment (at least, not in standard 3.5" formats).
>>
>>
>>> As for write endurance, you are about a decade out of touch...
>>
>>> <http://www.storagesearch.com/ssdmyths-endurance.html>
>>
>> Well, not for USB flash. I did recently torture a 2GB Kingston
>> USB to death and it had consistent data errors (with no error
>> meassege to make matter worse!) after about 3500 full overwrites.
>>

> USB flash devices are generally optimised for low costs rather than high
> quality or high endurance. They also often have very poor erase block
> management, since they have few chips and also must minimise the risk of
> data loss if the device is removed unexpectedly. This means a single
> block write to the device can cause many erase/writes to the flash.

>> Say the OP wants to overwrite his disk at 400MB/s, then 3500
>> full operwrites are reached after about 100 days of operation.
>>

> He doesn't say whether he wants to stream reads or writes, or how long
> he wants to sustain the transfers. My guess would be that he'd like to
> break off writing and do the occasional read - there are not many
> applications which produce new data at a rate of 34 TB per day, nothing
> of which needs to be kept for more than forty minutes!

I can think of an application or two that needs this. Quite
specialized though.

So here is a question to the OP: What is your access profile?

>> I expect SATA FLASH is better, but not all may be. What however
>> does not happen with modern FLASH is that writing a few 1000
>> times to a single location kills the drive. Traditional FLASH
>> without wear leveling had that problem, with some dying
>> after 10000...100000 writes to the same sector.
>>

> Modern SLC flash chips will have endurance in the range of at least a
> million erase/writes if you are nice to them (i.e., keep them at room
> temperature). MLC devices used in cheaper disks have significantly
> lower endurance.

> So for continuous writing using good SLC disks, he's got an average of
> something like 80 years before write endurance is a problem.

Yes, but even cheaper MLCs should be at 100'000 cycles today, that
is why I find the 3500 overwrites figure I found so disappointing,
and this thing uses an Intel flash memory chip, not some unpranded
or Tier-3 vendor product.

Expensive SLC is very hard to break and should indeed survive
decades at full write rate.

Arno
--
Arno Wagner, Dr. sc. techn., Dipl. Inform., CISSP -- Email: arno(a)wagner.name
GnuPG: ID: 1E25338F FP: 0C30 5782 9D93 F785 E79C 0296 797F 6B50 1E25 338F
----
Cuddly UI's are the manifestation of wishful thinking. -- Dylan Evans
From: Arno on
calypso(a)fly.srk.fer.hr.invalid wrote:
> Arno <me(a)privacy.net> kenjka:
>> Yes, but even cheaper MLCs should be at 100'000 cycles today, that

> MLC's often have at max 10.000 E/W cycles... SLC's have around 100.000...
> Reference is EMC's and STEC documentation for EFD drives (in fact it's STEC
> ZeusIOPS)...

>> Expensive SLC is very hard to break and should indeed survive
>> decades at full write rate.

> My calculations say that 400GB SLC drive can last 150 years with 24/7/365
> writes on it (SLC, 100.000 E/W cycles)...


Ok, say 400GB at 200MB/s. That gives 1.8 overwrites/h, i.e. about
55'000h, i.e. about 6.25 years. Sorry your math is off. And
the 6.25 years are with perfect wear leveling. Assuming my
experience of death after 3'500 cycles with 10'000 cycle MLC,
the disk could die as early as within 2 years.

Even 1'000'000 cycle MLC only gives you 20-60 years.

Arno
--
Arno Wagner, Dr. sc. techn., Dipl. Inform., CISSP -- Email: arno(a)wagner.name
GnuPG: ID: 1E25338F FP: 0C30 5782 9D93 F785 E79C 0296 797F 6B50 1E25 338F
----
Cuddly UI's are the manifestation of wishful thinking. -- Dylan Evans
From: Arno on
calypso(a)fly.srk.fer.hr.invalid wrote:
> Arno <me(a)privacy.net> kenjka:
>> > My calculations say that 400GB SLC drive can last 150 years with 24/7/365
>> > writes on it (SLC, 100.000 E/W cycles)...

>> Ok, say 400GB at 200MB/s. That gives 1.8 overwrites/h, i.e. about
>> 55'000h, i.e. about 6.25 years. Sorry your math is off. And
>> the 6.25 years are with perfect wear leveling. Assuming my
>> experience of death after 3'500 cycles with 10'000 cycle MLC,
>> the disk could die as early as within 2 years.

>> Even 1'000'000 cycle MLC only gives you 20-60 years.


> Why are you referring to SSD drive and sequential writes? The main reason
> why SSD are used are their high IOPS values! OK, I am talking from the
> storage vendor perspective, and not from the home-user perspective...

I am talking sustained maximum write speed. Does not need to be
sequential, but it is the worst-case for the lifetime. Of course
a lower rate with small writes that still result in an effective
write rate (because of larger internal block size) of 200MB/s
also hits this worst case.

And while the high IOP is one desirable parameter, it is not
the only one. For example an SSD can well be used for an external
filesystem journal. This is a mostly write and mostly sequential
write operation. However when you recover the journal the IOPs
are the bottleneck. So you may want to put your journal on the
SSD to bring recovery speed down dramatically. Or rollback
speed if it is a database journal.

For a home-user, OTOH, you may actually hit your figure. But
home users do not have 24/7 anyways.

So while your 150 years figure is certainly good to boost
sales, it is unusable to evaluate practical endurance. For
that you need to look at the particular worst case.

And there is a second problem. On power-fail a SSD can
corrupt areas not written to because of large internal
block sizes. That means in high-reliability applications
you actually can only write it in a sequential fashion
and without filesystem as everything else is dangerous to
your data.

The short summary is that SSDs have write issues that
you need to understand in order to decide whether to
use them or not. They shine on read IOPs though.

Arno
--
Arno Wagner, Dr. sc. techn., Dipl. Inform., CISSP -- Email: arno(a)wagner.name
GnuPG: ID: 1E25338F FP: 0C30 5782 9D93 F785 E79C 0296 797F 6B50 1E25 338F
----
Cuddly UI's are the manifestation of wishful thinking. -- Dylan Evans