A New Log-Structured Linux Caching Software Driver

[phoronix]

Phoronix: A New Log-Structured Linux Caching Software Driver

While it didn't make it as part of the initial staging pull for the Linux 3.12 merge window, on Sunday dm-writeboost was proposed for staging. DM-Writeboost is new log-structured Linux caching software...

http://www.phoronix.com/vr.php?view=MTQ1Mjg

[uid313]

Atomicity?

Does this negatively affect atomicity?
Atomic disk operations, like ACID compliance in databases.

[BradN]

My guess is it positively affects it, as writing out changes sequentially tends to let the software order things the right way without much effort. I'm imagining this works like an overgrown journal on ext3 or similar, but I don't really know exactly.

[uid313]

My guess is it positively affects it, as writing out changes sequentially tends to let the software order things the right way without much effort. I'm imagining this works like an overgrown journal on ext3 or similar, but I don't really know exactly.

But what if the software writes something to the disk, then later the power goes out.
Now the writes were never committed to the disk, they were in the cache and actually never written.

[Tobu]

But what if the software writes something to the disk, then later the power goes out.
Now the writes were never committed to the disk, they were in the cache and actually never written.

Disks use SCSI commands like FUA and FLUSH for barriers (ordering) and durability. Looking at the dm-writeboost pdf, it implements these commands by waiting for the SSD to return from a similar command. Barring bugs or exposure of the uncached device, filesystems will behave correctly. Reliability is lower since there are two points of failure (SSD and HDD), but you should be backing up regardless of how much hardware is involved.

[uid313]

Disks use SCSI commands like FUA and FLUSH for barriers (ordering) and durability. Looking at the dm-writeboost pdf, it implements these commands by waiting for the SSD to return from a similar command. Barring bugs or exposure of the uncached device, filesystems will behave correctly. Reliability is lower since there are two points of failure (SSD and HDD), but you should be backing up regardless of how much hardware is involved.

The thing about ACID and atomic operations is that they're supposed to be guaranteed.

[Tobu]

The thing about ACID and atomic operations is that they're supposed to be guaranteed.

You failed to read the short technical explanation, so here's a tl;dr. dm-writeboost doesn't change a thing for databases.