[olug] SSD write durability in production use

[Will Langford]

On Tue, Oct 14, 2008 at 9:16 AM, John Hobbs <john at velvetcache.org> wrote:

> Here's a post from Linus Torvalds about SSD's.  He's talking about how
> SSD's are fast on big chunk writes but terrible on small random
> writes.  He seems to really like his Intel SSD based on that.
>
>
> http://torvalds-family.blogspot.com/2008/10/so-i-got-one-of-new-intel-ssds.html
>
> - John Hobbs
>


This is a copy and paste from another email I received from a friend on a
different list.  It talks about MLC/SLC and other stuff... and given the
performance issue brought up by the above post... it seems relevant.  While
I didn't fully touch on the performance related stuffs for database IO on my
original post, I figured the implications were fairly obvious.

As a side note, on a test system... using a ram drive to hold the database
and working from that... didn't yield outstanding results.  Possible issues
would be version of the database (oldish), the hardware it was running on
(p4 celery 2ghz doesn't have the greatest memory access speeds and neutered
cache), and the fact that using system RAM prolly thrashed the memory
controller and similar... causing a huge bottle neck for an already-memory
sensitive database.

-Will

------ 8< snip 8< ------

If you use a better SSD, such as the Intel X25-E ESD devices, and you
should get much better life and performance. The X25E offers something
like 8000 4k IOPS in random-write mode, and much higher (20-30k I think,
don't have numbers handy) in sequential write.

The X25-E is a SLC (single level cell), which apparently gets around 10x
the write durability compared to a MLC (multi-level cell), which most
current SSDs (including the rest of the intel X25 range, the ones
without the -E denonmination) are.


http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3403  has some
interesting points.

The downside to the Intel SLCs are that they're small (32, 64, and maybe
128 GB drives out now), and relatively expensive, but for something like
a DB server, it should scream along.

Evgeniy ran some tests with a consumer grade MLC:
http://tservice.net.ru/~s0mbre/blog//devel/other/2008_08_20

Tobi Oetiker (MRTG/RRDtool guy) has some good points here too:
http://insights.oetiker.ch/linux/external-journal-on-ssd.html , although
he didn't benchmark it particularly robustly

------ 8< snip 8< ------