Winmodem-like solid state storage [LWN.net]

Winmodem-like solid state storage

Posted Apr 19, 2009 14:11 UTC (Sun) by oak (subscriber, #2786)
In reply to: Winmodem-like solid state storage by jzbiciak
Parent article: Linux Storage and Filesystem Workshop, day 2

> What I'd like to see is something I can get off the shelf at my local
Computer Mart (or on the web) that plugs into my PC and gives me raw
flash. Instead of focusing on "right sized" and "small" and "maximizing
battery life", it instead can be a bank of parallel flash such as what
Intel's SSD disks are, but with a raw interface. We can then use our
existing flash filesystems and infrastructure to drive those in a desktop
and laptop space

Unlike block based file systems like ext[234], the existing flash file
systems are designed for very small file systems. E.g. JFFS2 keeps the
whole file system metadata in RAM and is unusable in GB sized file
systems.

However, the newly merged UBIFS promises to work much better:
* http://lwn.net/Articles/275706/
* http://www.linux-mtd.infradead.org/doc/ubifs.html#L_scala...

There's not usage data on how well it performs with desktop and server
loads though.

(Log in to post comments)

Winmodem-like solid state storage

Posted Apr 19, 2009 14:28 UTC (Sun) by dwmw2 (subscriber, #2063) [Link]

" Unlike block based file systems like ext[234], the existing flash file systems are designed for very small file systems. E.g. JFFS2 keeps the whole file system metadata in RAM and is unusable in GB sized file systems."

Very true — although we put a lot of effort in to make JFFS2 better for OLPC with its 1GiB of NAND flash. It mounts in 6 seconds or so, and we reduced the RAM usage by a significant amount too. But still, JFFS2 was designed in the days of 32MiB or so of NOR flash, and definitely isn't intended to scale up to the kind of sizes we're seeing now.

UBIFS is much more promising, but as you correctly observe is not yet proven for desktop or server workloads. I'm actually keen to get btrfs working on raw flash, too.

The point is that with stuff done in software, we can do better; whether we do better or not today is a different, and less interesting issue.

After all, we can always implement the same "pretend to be a block device" kind of thing to tide us over in the short term, if we need to. We have three or four such translation layers in Linux already, and more on the way.