Re: [PATCH 00/21] np/pack-v4 updates

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, Sep 12, 2013 at 11:20 PM, Nicolas Pitre <nico@xxxxxxxxxxx> wrote:
> On Thu, 12 Sep 2013, Duy Nguyen wrote:
>
>> On Wed, Sep 11, 2013 at 11:25 PM, Nicolas Pitre <nico@xxxxxxxxxxx> wrote:
>> > On Wed, 11 Sep 2013, Duy Nguyen wrote:
>> >
>> >> Nico, if you have time you may want to look into this. The result v4
>> >> pack from pack-objects on git.git for me is 35MB (one branch) while
>> >> packv4-create produces 30MB (v2 is 40MB). I don't know why there is
>> >> such a big difference in size. I compared. Ident dict is identical.
>> >> Tree dict is a bit different (some that have same hits are ordered
>> >> differently). Delta chains do not differ much. Many groups of entries
>> >> in the pack are displaced though. I guess I turned a wrong knob or
>> >> something in pack-objects in v4 code..
>> >
>> > Will try to have a closer look.
>>
>> Problem found. I encoded some trees as ref-delta instead of pv4-tree
>> :( Something like this brings the size back to packv4-create output
>>
>> diff --git a/builtin/pack-objects.c b/builtin/pack-objects.c
>> index f604fa5..3d9ab0e 100644
>> --- a/builtin/pack-objects.c
>> +++ b/builtin/pack-objects.c
>> @@ -1490,7 +1490,8 @@ static void check_object(struct object_entry *entry)
>>   * deltify other objects against, in order to avoid
>>   * circular deltas.
>>   */
>> - entry->type = entry->in_pack_type;
>> + if (pack_version < 4)
>> + entry->type = entry->in_pack_type;
>>   entry->delta = base_entry;
>>   entry->delta_size = entry->size;
>>   entry->delta_sibling = base_entry->delta_child;
>
> Hmmm... I've folded this fix into your patch touching this area.

You may want to stricten the condition a bit, to "pack_version < 4 ||
entry->type != OBJ_TREE". I think always not doing it in v4 turns off
the reuse code path for blobs and tags.

> This code is becoming rather subtle and messy though.  We'll have to
> find a way to better abstract things.

Yep. Not sure how that should be done though. Maybe we can revisit it
when pack-objects learns to skip compatibility layer when reading v4
commits and trees..

> Especially since object data
> reuse will work only for blobs and tags with packv4.  Commits and trees
> will need adjustments to their indices.
-- 
Duy
--
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html




[Index of Archives]     [Linux Kernel Development]     [Gcc Help]     [IETF Annouce]     [DCCP]     [Netdev]     [Networking]     [Security]     [V4L]     [Bugtraq]     [Yosemite]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux SCSI]     [Fedora Users]