Mjk: first edit

2007-05-04T15:04:53Z

first edit

New page

== Introduction ==
Soft Updates increases file system efficiency by ordering the disk writes
so that every write creates a consistent state.

This allows all writes to happen asynchronously
and to avoid seeks wherever possible.

== File Dependencies ==
On a typical file system there are files, directories and pointers to these.
A directory inode points to the direcotry data
which points to either more directory nodes or to file inodes.
The file inodes point to the file data.

If this chain is broken somewhere in between,
for example if the directory points to inodes not (yet) written
and the system crashes,
it is in an inconsistent state.

== Ordered Writes ==
Soft Updates' Solution to this is ordering the meta-data writes to disk
to avoid such inconsistencies.

=== Dependency Tracking ===
To order these writes, the system needs to track which blocks
need to be written to disk before others.

Please consider the example
that one file is created an another is deleted,
both in the same directory.
(if you have problems imagining this,
please refer to the [http://www.hpl.hp.com/personal/Craig_Soules/papers/TOCS.softupdates.pdf fourth paper], page 8,
or click on [http://www2.informatik.hu-berlin.de/~kreikenb/su.pdf this link],
while it lasts).

If dependency information is stored regarding only blocks,
the directory block needs to be written before the inode block
and vice versa.

To solve this, the system remembers dependency information
for every single piece of data,
either file or directory.

Still, in the state from the example,
you can not write either block while guaranteeing consistency.

=== Roll-back ===
The solution to this is called roll-back:

One of the blocks is modified to a state that never existed
and then written to disk.
This extra write produces a sequences of consistent states

== Crash recovery ==
As a result,
whenever a crash occurs,
the whole file system is in a consistent state
in that no references point to some unused space.

There might be allocated blocks inside the system (mostly data)
that are not used anywhere.

These can be found by doing a check in the background.

== Efficiency ==
The speed increase in Soft Updates results
from writing all of the meta-data operations asynchronously.
In contrast with [[Journaling]],
no log is written to disk, producing no extra overhead.
The only extra writes that occur are the roll-back writes
and those resulting when the current state does not fit to memory any more.

In tests in the [http://www.usenix.org/publications/library/proceedings/usenix2000/general/full_papers/seltzer/seltzer.pdf literature],
SoftUpdates results in a speed increase of 30 to 660 %.

== Links ==
*[http://www.usenix.org/publications/library/proceedings/usenix2000/general/full_papers/seltzer/seltzer.pdf Journaling versus Soft Updates: Asynchronous Meta-data Protection in File Systems]
*[http://pdos.csail.mit.edu/papers/cffs-usenix97.ps Embedded Inodes and Explicit Grouping: Exploiting Disk Bandwidth for Small files]
*[http://pdos.csail.mit.edu/6.824/papers/hagmann-fs.pdf Reimplementing the Cedar File System Using Logging and Group Commit]
*[http://www.hpl.hp.com/personal/Craig_Soules/papers/TOCS.softupdates.pdf Soft Updates: A Solution to the Metadata Update Problem in File Systems]
* http://en.wikipedia.org/wiki/Comparison_of_file_systems

Soft-Updates - Revision history

Mjk: first edit