Re: [Gnu-arch-users] ANNOUNCEMENT -- "timestamps" optimization

[Miles Bader]

I'm seeing similar numbers with the emacs source tree -- a little bit better
than 2x, about 30s for `what-changed' on a hopefully uncached tree,
w/,,inode-sigs vs. 1m15s w/o.  [It's a bit hard to tell for sure, since I
don't know a reliable and convenient method to flush the disk cache --
starting lots of big apps multiple times is a pain, and just touching lots of
mallocated memory in a program doesn't seem to work reliably.]  It seems to
speed up even memory-cached trees by about 2x as well.

One obvious improvement might be to store the filename as well as the file's
id-tag in ,,inode-sigs.  Especially for taglines, if you know the file's not
changed, then you know the tagline hasn't changed either, and I think reading
even parts of files to get the tagline can result in plenty of disk seeking
(whereas inodes are small and often packed into disk blocks together).  Even
for explicit tags, you can use the ,,inode-sigs info to see if the explicit
tag file has changed (hmmm, there's one advantage of using individual .id
files!).

IOW, make the inventory code read ,,inode-sigs too, and use it to avoid
reading the first/last 1K of the file-contents for taglines, or .id files for
explicit tags (except for newly-added or touched files).

-Miles
-- 
`Suppose Korea goes to the World Cup final against Japan and wins,' Moon said.
`All the past could be forgiven.'   [NYT]