git/list[1] front-page[2] threads[3] people[4] search[5] about
 

[PATCH 0/3] low-hanging performance fruit with promisor packs

From
Jeff King <peff@peff.net>
Date
Apr 13, 2021, 07:12 UTC
Message-ID
<YHVECXHfZ1bidTJH@coredump.intra.peff.net>
In-Reply-To
<YHTcHY+P7RuZJGab@coredump.intra.peff.net>
On Mon, Apr 12, 2021 at 07:47:41PM -0400, Jeff King wrote:
> The patch below drops the peak heap to 165MB. Still quite a bit more,
> but I think it's a combination of delta-base cache (96MB) plus extra
> structs for all the non-commit objects whose flags we marked.

I think we can do even better than that, after looking into the "do we really need to parse the objects?" comment I left (spoiler: the answer is no, we do not need to, at least for that caller).

Here are some cleaned-up patches that I think improve the situation quite a bit. This is just the low-hanging fruit from this part of the discussion; I'm sure there's more to do to make using partial clones pleasant. In particular:

  - this does nothing for the "oops, we turned all of the promisor
    objects loose and then deleted them" problem. Hopefully Rafael
    will produce a nice patch for that
  - In is_promisor_object(), we still call parse_object(), because it
    really does look at the contents. But doing so for blobs is wasteful
    (it's a lot of bytes we push through sha1, and we don't even look at
    them). It might be worth using oid_object_info() to avoid this. This
    introduces a little overhead, but I think would be a net win (it
    would be really nice if we could amortize the object lookup work;
    i.e., if there was a way to call oid_object_info_extended() and say
    "open the object and look at the type; only return the contents if
    it's a non-blob").
  - I still think it's probably worth having a mode where we store the
    set of pointed-to objects we don't have, rather than parsing on the
    fly. This could easily be made optional (it's a good thing if you
    _have_ a lot of objects that point to a small or moderate number of
    missing objects, like a blob:limit or blob:none filter; it's a bad
    thing if you have very few objects but they point to a very large
    number).

I didn't explore any of those here, and I don't plan to look into them anytime soon. I'm just documenting my findings for later.

Anyway, here are the patches.
  [1/3]: is_promisor_object(): free tree buffer after parsing
  [2/3]: lookup_unknown_object(): take a repository argument
  [3/3]: revision: avoid parsing with --exclude-promisor-objects
 builtin/fsck.c                   |  2 +-
 builtin/pack-objects.c           |  2 +-
 http-push.c                      |  2 +-
 object.c                         |  7 +++----
 object.h                         |  2 +-
 packfile.c                       |  1 +
 refs.c                           |  2 +-
 revision.c                       |  2 +-
 t/helper/test-example-decorate.c |  6 +++---
 t/perf/p5600-partial-clone.sh    | 12 ++++++++++++
 upload-pack.c                    |  2 +-
 walker.c                         |  2 +-
 12 files changed, 27 insertions(+), 15 deletions(-)
-Peff
Previous: Jeff KingNext: Jeff King
Message 11 of 46 in “rather slow 'git repack' in 'blob:none' partial clones”
  1. SZEDER GáborApr 3, 2021
  2. Rafael SilvaApr 5, 2021
  3. Jeff KingApr 7, 2021
  4. Jonathan TanApr 8, 2021
  5. Jeff KingApr 8, 2021
  6. Rafael SilvaApr 12, 2021
  7. SZEDER GáborApr 12, 2021
  8. Bryan TurnerApr 12, 2021
  9. Jeff KingApr 12, 2021
  10. Jeff KingApr 12, 2021
  11. 0/3 low-hanging performance fruit with promisor packsJeff King, Apr 13, 2021
  12. 1/3 is_promisor_object(): free tree buffer after parsingJeff King, Apr 13, 2021
  13. Junio C HamanoApr 13, 2021
  14. Jeff KingApr 14, 2021
  15. 2/3 lookup_unknown_object(): take a repository argumentJeff King, Apr 13, 2021
  16. 3/3 revision: avoid parsing with --exclude-promisor-objectsJeff King, Apr 13, 2021
  17. Junio C HamanoApr 13, 2021
  18. SZEDER GáborApr 13, 2021
  19. Jonathan TanApr 14, 2021
  20. Rafael SilvaApr 14, 2021
  21. SZEDER GáborApr 13, 2021
  22. Jeff KingApr 14, 2021
  23. SZEDER GáborApr 11, 2021
  24. Rafael SilvaApr 12, 2021
  25. 0/2 prevent `repack` to unpack and delete promisor objectsRafael Silva, Apr 14, 2021
  26. 1/2 repack: teach --no-prune-packed to skip `git prune-packed`Rafael Silva, Apr 14, 2021
  27. Jonathan TanApr 14, 2021
  28. Rafael SilvaApr 18, 2021
  29. 2/2 repack: avoid loosening promisor pack objects in partial clonesRafael Silva, Apr 14, 2021
  30. Jonathan TanApr 15, 2021
  31. Junio C HamanoApr 15, 2021
  32. Jeff KingApr 15, 2021
  33. Jeff KingApr 15, 2021
  34. Rafael SilvaApr 18, 2021
  35. Junio C HamanoApr 15, 2021
  36. Rafael SilvaApr 18, 2021
  37. Junio C HamanoApr 14, 2021
  38. Jeff KingApr 15, 2021
  39. Rafael SilvaApr 18, 2021
  40. 0/1 prevent `repack` to unpack and delete promisor objectsRafael Silva, Apr 18, 2021
  41. 1/1 repack: avoid loosening promisor objects in partial clonesRafael Silva, Apr 18, 2021
  42. Jonathan TanApr 19, 2021
  43. Rafael SilvaApr 21, 2021
  44. Junio C HamanoApr 19, 2021
  45. Rafael SilvaApr 21, 2021
  46. repack: avoid loosening promisor objects in partial clonesRafael Silva, Apr 21, 2021

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.