Re: Add note about IANA confirmation

From: Linus Torvalds <torvalds@osdl.org>
Date: Sep 12, 2005, 22:40 UTC
Message-ID: <Pine.LNX.4.58.0509121523170.3266@g5.osdl.org>
In-Reply-To: <7vk6hm7ymd.fsf@assigned-by-dhcp.cox.net>

On Mon, 12 Sep 2005, Junio C Hamano wrote:

Show 5 quoted lines

> 
> I noticed one thing that is a bit curious.  Packing the same set
> of objects with --window=50 seems to give worse packing result
> (space-wise) than with --window=10, which is the default if I
> recall correctly.

The window doesn't seem to make much of a difference any more, now that we sort object blobs by filename. I suspect that it's just being unlucky: it finds a locally smaller diff and uses it, but then that ends up causing some other delta to _not_ be generated, because that would overflow the "--depth".

At least I _think_ that's the case. Maybe there's something really wrong in the window handling.

Generally, it's cheaper and more efficient to just increase the "depth" rather than the "window". Of course, that has a possibly very bad run-time overhead at unpacking time, so increasing it too much is bad.

Quick test (maybe I did it right, maybe I didn't):

			window=10	window=50

	 depth=10	 57240724	 57411879
	 depth=50	 56700053	 56808913
	 depth=inf	 56603101	 56541663

ie "--window=50" is actually worse than "--window=10" for depths 10 and 50, but better for the "infinite" depth case (approximated by using "--depth=50000" which is practically infinite, since I just packed the 2.6.12 history with less than 50,000 objects ;)

Of course, the 2.6.12 tree is not a good pack object. Half the objects in that tree have no history at all. It's not clear that it's particularly representative of any real long-term history.

		Linus