On Sun, May 06, 2001 at 11:27:11AM -0400, Patrick J. LoPresti wrote:
> Keith Matthews <keith_m@xxxxxxxxxxxxxxxxxxx> writes:
> > The other difference from what I have read is that ext3 allows journalling
> > of the file content data as well as the metadata, the other three (as far
> > as I can see) only journal metadata. This may be important to you.
> I apologize if this is a stupid question, but how could data
> journalling possibly matter to anybody?
> I mean, as long as the data are written to disk before the metadata
> are written to the journal, there is no reliability to be gained by
> journalling the data. What am I missing?
Performance. There are certain applications, including mail and NFS
serving, where processes want data written synchronously to disk in
very small chunks (the size of a page in NFSv2, or the size of an
email for sendmail). That translates to a _lot_ of disk seeks.
If you can do the synchronous writes to a single journal instead, then
the synchronous part of the filesystem update becomes a matter of
streaming sequentially to a log file instead (and the filesystem can
do its normal writeback caching when updating the primary copy of the
files, reordering those writes to go as quickly as possible in the
background at some later point). Put the log on a separate disk from
the main filesystem and it goes even faster.