[olug] SSD write durability in production use

[Will Langford]

We've been toying with the thought of SSD drives for database usage here at
work.  Insanely fast IO access seen as the primary reason to try it out.
 Write durability questions came up... and there was a fair bit of debate
over it, without actually giving any numbers other than "i think....".

http://wiki.eeeuser.com/ssd_write_limit

This is someone's math response to SSD write durability on EEE.  It takes
into account a sub-industry-standard wear leveling and stuff... and in
general seem so to be fairly decent.

The only flaw I've found with the analysis is that it assumes that there is
no static content on the drive.  OS, libraries, etc.  Stuff that doesn't
move or rewrite and therefore detracts from the wear leveling stuff.

Now, onto the solution for stuff I came up with prior to finding this
article, which somewhat mitigates the flaw in the math.  The idea was to buy
a SSD that was 2x or 4x the size of what you were planning on using (given
small size of DB's and price breaks on drives -- this isn't infeasible).
 Then make the partition you'll actively use be only that fraction bit of
the total size of the drive.  This way, you'd assume the wear leveling will
extend the writes to the rest of the drive.  Big assumption.  Assuming it
does indeed work that way, you effectively take your write cycle count and
multiply it by the reciprocal fractional size of your partition.

Given the chart, at 1MB/sec writing at 50% wear leveling optimization
stuffs, you'd get 6 years of life off the drive.  Possibily adjust this by
the fractional math and you have an interesting prospect.

-Will