Re: [gentoo-portage-dev] Re: search functionality in emerge

[Alec Warner]

On Mon, Dec 1, 2008 at 4:20 PM, Tambet <qtvali@×××××.com> wrote:
> 2008/12/2 Emma Strubell <emma.strubell@×××××.com>
>>
>> True, true. Like I said, I don't really use overlays, so excuse my
>> igonrance.
>
> Do you know an order of doing things:
>
> Rules of Optimization:
>
> Rule 1: Don't do it.
> Rule 2 (for experts only): Don't do it yet.
>
> What this actually means - functionality comes first. Readability comes
> next. Optimization comes last. Unless you are creating a fancy 3D engine for
> kung fu game.
>
> If you are going to exclude overlays, you are removing functionality - and,
> indeed, absolutely has-to-be-there functionality, because noone would
> intuitively expect search function to search only one subset of packages,
> however reasonable this subset would be. So, you can't, just can't, add this
> package into portage base - you could write just another external search
> package for portage.
>
> I looked this code a bit and:
> Portage's "__init__.py" contains comment "# search functionality". After
> this comment, there is a nice and simple search class.
> It also contains method "def action_sync(...)", which contains
> synchronization stuff.
>
> Now, search class will be initialized by setting up 3 databases - porttree,
> bintree and vartree, whatever those are. Those will be in self._dbs array
> and porttree will be in self._portdb.
>
> It contains some more methods:
> _findname(...) will return result of self._portdb.findname(...) with same
> parameters or None if it does not exist.
> Other methods will do similar things - map one or another method.
> execute will do the real search...
> Now - "for package in self.portdb.cp_all()" is important here ...it
> currently loops over whole portage tree. All kinds of matching will be done
> inside.
> self.portdb obviously points to porttree.py (unless it points to fake tree).
> cp_all will take all porttrees and do simple file search inside. This method
> should contain optional index search.
>
> 		self.porttrees = [self.porttree_root] + \
> 			[os.path.realpath(t) for t in self.mysettings["PORTDIR_OVERLAY"].split()]
>
> So, self.porttrees contains list of trees - first of them is root, others
> are overlays.
>
> Now, what you have to do will not be harder just because of having overlay
> search, too.
>
> You have to create method def cp_index(self), which will return dictionary
> containing package names as keys. For oroot... will be "self.porttrees[1:]",
> not "self.porttrees" - this will only search overlays. d = {} will be
> replaced with d = self.cp_index(). If index is not there, old version will
> be used (thus, you have to make internal porttrees variable, which contains
> all or all except first).
>
> Other methods used by search are xmatch and aux_get - first used several
> times and last one used to get description. You have to cache results of
> those specific queries and make them use your cache - as you can see, those
> parts of portage are already able to use overlays. Thus, you have to put
> your code again in beginning of those functions - create index_xmatch and
> index_aux_get methods, then make those methods use them and return their
> results unless those are None (or something other in case none is already
> legal result) - if they return None, old code will be run and do it's job.
> If index is not created, result is None. In index_** methods, just check if
> query is what you can answer and if it is, then answer it.
>
> Obviously, the simplest way to create your index is to delete index, then
> use those same methods to query for all nessecary information - and fastest
> way would be to add updating index directly into sync, which you could do
> later.
>
> Please, also, make those commands to turn index on and off (last one should
> also delete it to save disk space). Default should be off until it's fast,
> small and reliable. Also notice that if index is kept on hard drive, it
> might be faster if it's compressed (gz, for example) - decompressing takes
> less time and more processing power than reading it fully out.

I'm pretty sure your mistaken here, unless your index is stored on a
floppy or something really slow.

A disk read has 2 primary costs.

Seek Time: Time for the head to seek to the sector of disk you want.
Spin Time: Time for the platter to spin around such that the sector
you want is under the read head.

Spin Time is based on rpm, so average 7200 rpm / 60 seconds = 120
rotations per second, so worst case (you just passed the sector you
need) you need to wait 1/120th of a second (or 8ms).

Seek Time is per hard drive, but most drives provide average seek
times under 10ms.

So it takes on average 18ms to get to your data, then you start
reading.  The index will not be that large (my esearchdb is 2 megs,
but lets assume 10MB for this compressed index).

I took a 10MB meg sqlite database and compressed it with gzip (default
settings) down to 5 megs.
gzip -d on the database takes 300ms, catting the decompressed data
base takes 88ms (average of 5 runs, drop disk caches between runs).

I then tried my vdb_metadata.pickle from /var/cache/edb/vdb_metadata.pickle

1.3megs compresses to 390k.

36ms to decompress the 390k file, but 26ms to read the 1.3meg file from disk.

Your index would have to be very large or very fragmented on disk
(requiring more than one seek) to see a significant gain in file
compression (gzip scales linearly).

In short, don't compress the index ;p

>
> Have luck!
>
>>> -----BEGIN PGP SIGNED MESSAGE-----
>>> Hash: SHA1
>>>
>>> Emma Strubell schrieb:
>>> > 2) does anyone really need to search an overlay anyway?
>>>
>>> Of course. Take large (semi-)official overlays like sunrise. They can
>>> easily be seen as a second portage tree.
>>> -----BEGIN PGP SIGNATURE-----
>>> Version: GnuPG v2.0.9 (GNU/Linux)
>>> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org
>>>
>>> iEYEARECAAYFAkk0YpEACgkQ4UOg/zhYFuD3jQCdG/ChDmyOncpgUKeMuqDxD1Tt
>>> 0mwAn2FXskdEAyFlmE8shUJy7WlhHr4S
>>> =+lCO
>>> -----END PGP SIGNATURE-----
>>>
>> On Mon, Dec 1, 2008 at 5:17 PM, René 'Necoro' Neumann <lists@××××××.eu>
>> wrote:
>>
>
>