[OpenIndiana-discuss] Good SLOG devices?

[wessels]

Go for the ups AND a device which doesn't lose data. Not only the loss
of power to the machine can make the slog lose it's data when it isn't
internally protected by a battery.
I think you can estimating how much gets written to your ideal slog
can be measured like this:
-turn the zil off completely. measure performance and how data gets
written in a certain period.
-turn the zil on again. repeat the same workload as before measure
performance and much zilstat records in a certain time.
The difference in performance is the best you can expect.
Multiply the amount of data written to the zil by the above factor and
you have a reasonable indication how much gets written to your slog
per day.

Good news last night while we were discussing this a very interesting
post appeared on the zfs mailinglist
http://mail.opensolaris.org/pipermail/zfs-discuss/2011-March/047613.html
What appears the world's first dram based disk in 3.5" form factor.
Looks very interesting!
Pro's
-specifically engineered for this workload.
-has a SAS interface, HA folks would be very happy about this
-made by a company which knows how to build enterprise stuff
The only negative side I can see is the price, but I don't know the pricing.
And it doesn't fit in your x4150... but if you really want this you
can as well buy a small SuperMicro chassis with 3.5" bays as JBOD and
a SAS card for the x4150. Gives you even more expension.


On Wed, Mar 2, 2011 at 8:52 AM, Ivar Janmaat <ijanmaat at xs4all.nl> wrote:
> Thank you!
>
> I read in the DDRdrive paper that X-25E does not have  a power cell onboard
> to complete operations if power fails.
> The Vertex 2 Pro does have some capacitors to complete operations after
> power failure.
> I think that makes the Vertex a bit more interesting. Or should I go for a
> UPS?
> I have already done some statistics with zilstate. I will try to see if I
> can also do some wear calculations.
> The problem is however that I first need to build the system so it can
> support lots of VMs.
> The proper ZIL statistics will be only then available when the system
> reaches full load.
>
> Kind regards,
>
> Ivar
>
> wessels schreef:
>>
>> forget the acard for the x4150
>>
>> there aren't many datapoints (experiences, tests) about the OCZ Vertex
>> 2 series, not like the intel SSD's or the DDRdrive and I've no
>> experience with them or any SSD as SLOG for that matter.
>> Yes the zil is checksummoned. But you need another copy of the data.
>> Otherwise when the zil is read, and errors are detected, from the slog
>> device you're still in trouble. Remember the zil is rarely read, it's
>> mainly write-only. But when the slog devices fails during normal
>> operation it's offlined and the zil is written on the pool like
>> without the slog.
>> The mtbf isn't correlated with the wear of the device. Every time you
>> (implicit) erase a page the cells in the page wear a bit. As rule of
>> thumb SLC cells can be programmed 100000x and MLC cells 10000x. So
>> after these erases the cell's can't be used anymore. 10000 writes is a
>> bit short but the SSD employ wear-leveling techniques to ensure that
>> all cells gets evenly erased. How long your SSD last is a function of
>> how much data you write to it. A smaller SSD will wear out earlier
>> than a larger SSD of the same product line because of wear leveling.
>> Google the calculations. And newer is not always better. The new 25nm
>> generation of flash memory wears out much faster than previous 34nm
>> generation.
>> To determine how much the zil is used you can use the excellent tool
>> from Richard Elling at
>> http://www.richardelling.com/Home/scripts-and-programs-1/zilstat By
>> monitoring how much your zil is used you can determine how much is
>> written to it in say a day. With that number you can calculate when
>> your ssd wears out.
>> TRIM is not needed on good SSD. Look at early intel SSD's they didn't
>> have TRIM and still performed excellent. Remember the zil is very
>> small a couple of gigabytes at the most. So the rest of the drive can
>> be used for wear-leveling. Maybe the drive can be helped be limiting
>> it's capacity via e.g. DCO you only need 4 or 6GB for the zil. But you
>> need to test this.
>> All of the above is explained in much more depth in the ddrdrive paper
>> I mentioned earlier.
>>
>> The DDRdrive and the Intel X25E are validated solutions all other
>> drives are a gamble...
>> Labeling a product "enterprise" doesn't necessary mean it's a
>> validated enterprise product...
>>
>> Good luck
>>
>>
>
>
> _______________________________________________
> OpenIndiana-discuss mailing list
> OpenIndiana-discuss at openindiana.org
> http://openindiana.org/mailman/listinfo/openindiana-discuss
>