Re: [RFC] Optimize diff-delta.c

On 2007-05-01 16:05:24, Nicolas Pitre wrote:

quoted

On Tue, 1 May 2007, Martin Koegler wrote:

Right.  I think it would be a good idea to extend the delta format as 
well to allow for larger offsets in pack v4.

Is git://repo.or.cz/git/fastimport.git#sp/pack4 the current version of
pack v4 efforts?

quoted

The delta index has approximately the same size in memory as the
uncompressed blob ((blob size)/16*(sizeof(index_entry)).

One thing that could be done with really large blobs is to create a 
sparser index, i.e. have a larger step than 16.  Because the delta match 
loop scans backward after a match the sparse index shouldn't affect 
compression that much on large blobs and the index could be 
significantly smaller.

In the long term, I think, that the delta generation code needs to get
tunable.

I would add a init_delta function for reading the configuration
options. In create_delta and create_delta_index, different delta
heuristics can be selected based on the blob size and the
configuration options. As patch_delta is not affected by this, it
should be easy to integrate new stragegies.

quoted

I tried to speed up the delta generation by searching for a common 
prefix, as my blobs are mostly append only. I tested it with about 
less than 1000 big blobs. The time for finding the deltas decreased 
from 17 to 14 minutes cpu time.

I'm surprised that your patch makes so much of a difference.  Normally 
the first window should always match in the case you're trying to 
optimize and the current code should already perform more or less the 
same as your common prefix match does.

A block is limited to 64k. If the file has some hundred MBs, it has to
match many blocks.

My patch can process everything except the few last thousand lines by
doing a memcmp.

Additionally, nearly every line starts with the same, longer than 16
byte prefix. So its likely, that many blocks map to the same hash
value.

Ah, no, actually what your patch does is a pessimisation of the matching 
code by not considering other and possibly better matches elsewhere in 
the reference buffer whenever there is a match at the beginning of both 
buffers.  I don't think this is a good idea in general.

For small files, I agree with you.

What you should try instead if you want to make the process faster is to 
lower the treshold used to consider a match sufficiently large to stop 
searching.  That has the potential for even faster processing as the 
"optimization" would then be effective throughout the buffer and not 
only at the beginning.

[...]

You could experiment with that value to determine the best speed vs size 
compromize.

I will do some experiments.

mfg Martin Kögler
PS: Sorry for break threading, as you did not CC me.

On Fri, 4 May 2007, Martin Koegler wrote:

On 2007-05-01 16:05:24, Nicolas Pitre wrote:

quoted

quoted

On Tue, 1 May 2007, Martin Koegler wrote:

Right.  I think it would be a good idea to extend the delta format as 
well to allow for larger offsets in pack v4.

Is git://repo.or.cz/git/fastimport.git#sp/pack4 the current version of
pack v4 efforts?

Yes.  It is lagging behind current git though, and not usable yet.

quoted

quoted

The delta index has approximately the same size in memory as the
uncompressed blob ((blob size)/16*(sizeof(index_entry)).

One thing that could be done with really large blobs is to create a 
sparser index, i.e. have a larger step than 16.  Because the delta match 
loop scans backward after a match the sparse index shouldn't affect 
compression that much on large blobs and the index could be 
significantly smaller.

In the long term, I think, that the delta generation code needs to get
tunable.

No.  It should be self-tunable certainly, but there are way too many 
config options already, and another one for the inner working of the 
delta algorithm would be a bit too esoteric for most users and they 
won't get advantage of it.  This thing really has to self tune itself.

quoted

quoted

I tried to speed up the delta generation by searching for a common 
prefix, as my blobs are mostly append only. I tested it with about 
less than 1000 big blobs. The time for finding the deltas decreased 
from 17 to 14 minutes cpu time.

I'm surprised that your patch makes so much of a difference.  Normally 
the first window should always match in the case you're trying to 
optimize and the current code should already perform more or less the 
same as your common prefix match does.

A block is limited to 64k. If the file has some hundred MBs, it has to
match many blocks.

Only if the first match is smaller than 64K.  If the first match is 64K 
in size then the rest of the file is not considered at all.

My patch can process everything except the few last thousand lines by
doing a memcmp.

Additionally, nearly every line starts with the same, longer than 16
byte prefix. So its likely, that many blocks map to the same hash
value.

The hash index only remembers the lowest of consecutive identical blocks 
so repeated blocks are indexed only with the first one.  If however you 
happen to have many identical blocks interlaced between other blocks 
then there is not much that can be done.  What the code does in that 
case is to trim those hash buckets that gets too large by keeping only a 
few entries across the reference buffer to avoid a O(n^2) behavior. But 
that happens only when your line beginnings are located on the same 
block boundary (but with a block size of 16 this is rather likely in the 
presence of lots of lines I suppose).

I'll be very interested in the results you get with my suggested patch.


Nicolas