Re: [tlug] SSD filesystem

[Francois Cartegnie <fcartegnie@example.com>]

Le lundi 8 juin 2009, Edward Middleton a écrit :

> When you talk about SSD, usb flash, CF, etc. you aren't talking about
> direct flash access.  There is no guarantee that any operations on the
> device will actually translate into any specific operation on the
> flash.  Trying to second guess the controller on an immature device like
> SSD doesn't make sense.

Hi Edward,

The particular case of SD flash cards is that most cheap cards internal 
controllers (as SD does only half the job) doesn't do wear levelling, while 
other does (and optimized for FAT), and you have absolutely no way to know 
which ones.

> If you are directly accessing the flash then it makes sense to consider
> wear-leveling etc.  You can do this though the MTD[1] interface using a
> filesystem like jffs2[2] now.

Even if jffs if optimized for flash chips and used in embedded devices, it's 
usually only for storing configuration or extensions. Systems designers avoid 
putting pressure on the flash, because they have no way to have information 
of wear levels.

Now that SMART extensions have done it in the industry, i don't see why we 
wouldn't see something like for flash based devices. Wear stats, reallocation 
count, available free blocks for reallocation and prefail values are a must 
for trusting those storage devices (especially when lifetime is estimated 
with a non journalized filesystem).

Francois