git@vger.kernel.org list mirror (unofficial, one of many)
 help / color / mirror / code / Atom feed
From: Brandon Casey <drafnel@gmail.com>
To: Jeff King <peff@peff.net>
Cc: "git@vger.kernel.org" <git@vger.kernel.org>
Subject: Re: [PATCH 4/4] pack-revindex: radix-sort the revindex
Date: Mon, 8 Jul 2013 13:50:41 -0700	[thread overview]
Message-ID: <CA+sFfMfQcnx+OGNd+v7NJC5zSXg2OR1QiLyRSqDjXD0zb4mvtA@mail.gmail.com> (raw)
In-Reply-To: <20130707101438.GD19143@sigill.intra.peff.net>

On Sun, Jul 7, 2013 at 3:14 AM, Jeff King <peff@peff.net> wrote:
> The pack revindex stores the offsets of the objects in the
> pack in sorted order, allowing us to easily find the on-disk
> size of each object. To compute it, we populate an array
> with the offsets from the sha1-sorted idx file, and then use
> qsort to order it by offsets.
>
> That does O(n log n) offset comparisons, and profiling shows
> that we spend most of our time in cmp_offset. However, since
> we are sorting on a simple off_t, we can use numeric sorts
> that perform better. A radix sort can run in O(k*n), where k
> is the number of "digits" in our number. For a 64-bit off_t,
> using 16-bit "digits" gives us k=4.
>
> On the linux.git repo, with about 3M objects to sort, this
> yields a 400% speedup. Here are the best-of-five numbers for
> running "echo HEAD | git cat-file --batch-disk-size", which
> is dominated by time spent building the pack revindex:
>
>           before     after
>   real    0m0.834s   0m0.204s
>   user    0m0.788s   0m0.164s
>   sys     0m0.040s   0m0.036s
>
> On a smaller repo, the radix sort would not be
> as impressive (and could even be worse), as we are trading
> the log(n) factor for the k=4 of the radix sort. However,
> even on git.git, with 173K objects, it shows some
> improvement:
>
>           before     after
>   real    0m0.046s   0m0.017s
>   user    0m0.036s   0m0.012s
>   sys     0m0.008s   0m0.000s
>
> Signed-off-by: Jeff King <peff@peff.net>
> ---
> I think there are probably still two potential issues here:
>
>   1. My while() loop termination probably has issues when we have to use
>      all 64 bits to represent the pack offset (not likely, but...)
>
>   2. We put "int pos[65536]" on the stack. This is a little big, but is
>      probably OK, as I think the usual small stack problems we have seen
>      are always in threaded code. But it would not be a big deal to heap
>      allocate it (it would happen once per radix step, which is only 4
>      times for the whole sort).
>
>  pack-revindex.c | 77 +++++++++++++++++++++++++++++++++++++++++++++++++++++----
>  1 file changed, 72 insertions(+), 5 deletions(-)
>
> diff --git a/pack-revindex.c b/pack-revindex.c
> index 77a0465..d2adf36 100644
> --- a/pack-revindex.c
> +++ b/pack-revindex.c
> @@ -59,11 +59,78 @@ static int cmp_offset(const void *a_, const void *b_)
>         /* revindex elements are lazily initialized */
>  }
>
> -static int cmp_offset(const void *a_, const void *b_)
> +/*
> + * This is a least-significant-digit radix sort using a 16-bit "digit".
> + */
> +static void sort_revindex(struct revindex_entry *entries, int n, off_t max)

If 'n' is the number of objects in the pack, shouldn't it be unsigned?

The data type for struct packed_git.num_objects is uint32_t.  Looks
like create_pack_revindex uses the wrong datatype when it captures
num_objects in the int num_ent and passes it to sort_revindex.  So, it
looks like that function needs to be updated too.

>  {
> -       const struct revindex_entry *a = a_;
> -       const struct revindex_entry *b = b_;
> -       return (a->offset < b->offset) ? -1 : (a->offset > b->offset) ? 1 : 0;
> +       /*
> +        * We need O(n) temporary storage, so we sort back and forth between
> +        * the real array and our tmp storage. To keep them straight, we always
> +        * sort from "a" into buckets in "b".
> +        */
> +       struct revindex_entry *tmp = xcalloc(n, sizeof(*tmp));
> +       struct revindex_entry *a = entries, *b = tmp;
> +       int digits = 0;
> +
> +       /*
> +        * We want to know the bucket that a[i] will go into when we are using
> +        * the digit that is N bits from the (least significant) end.
> +        */
> +#define BUCKET_FOR(a, i, digits) ((a[i].offset >> digits) & 0xffff)
> +
> +       while (max / (((off_t)1) << digits)) {

Is there any reason this shouldn't be simplified to just:

       while (max >> digits) {

I glanced briefly at the assembly and it appears that gcc does
actually emit a divide instruction to accomplish this, which I think
we can avoid by just rearranging the operation.

> +               struct revindex_entry *swap;
> +               int i;
> +               int pos[65536] = {0};
> +
> +               /*
> +                * We want pos[i] to store the index of the last element that
> +                * will go in bucket "i" (actually one past the last element).
> +                * To do this, we first count the items that will go in each
> +                * bucket, which gives us a relative offset from the last
> +                * bucket. We can then cumulatively add the index from the
> +                * previous bucket to get the true index.
> +                */
> +               for (i = 0; i < n; i++)
> +                       pos[BUCKET_FOR(a, i, digits)]++;
> +               for (i = 1; i < ARRAY_SIZE(pos); i++)
> +                       pos[i] += pos[i-1];
> +
> +               /*
> +                * Now we can drop the elements into their correct buckets (in
> +                * our temporary array).  We iterate the pos counter backwards
> +                * to avoid using an extra index to count up. And since we are
> +                * going backwards there, we must also go backwards through the
> +                * array itself, to keep the sort stable.
> +                */
> +               for (i = n - 1; i >= 0; i--)
> +                       b[--pos[BUCKET_FOR(a, i, digits)]] = a[i];
> +
> +               /*
> +                * Now "b" contains the most sorted list, so we swap "a" and
> +                * "b" for the next iteration.
> +                */
> +               swap = a;
> +               a = b;
> +               b = swap;
> +
> +               /* And bump our digits for the next round. */
> +               digits += 16;
> +       }
> +
> +       /*
> +        * If we ended with our data in the original array, great. If not,
> +        * we have to move it back from the temporary storage.
> +        */
> +       if (a != entries) {
> +               int i;
> +               for (i = 0; i < n; i++)
> +                       entries[i] = tmp[i];

I think I recall that somebody investigated whether a for loop like
you have above was faster for copying structures than memcpy.  I
forget whether it was conclusive.  Did you happen to compare them?

<snip>

-Brandon

  parent reply	other threads:[~2013-07-08 20:50 UTC|newest]

Thread overview: 52+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2013-07-07 10:01 [RFC/PATCH 0/4] cat-file --batch-disk-sizes Jeff King
2013-07-07 10:03 ` [PATCH 1/4] zero-initialize object_info structs Jeff King
2013-07-07 17:34   ` Junio C Hamano
2013-07-07 10:04 ` [PATCH 2/4] teach sha1_object_info_extended a "disk_size" query Jeff King
2013-07-07 10:09 ` [PATCH 3/4] cat-file: add --batch-disk-sizes option Jeff King
2013-07-07 17:49   ` Junio C Hamano
2013-07-07 18:19     ` Jeff King
2013-07-08 11:04     ` Duy Nguyen
2013-07-08 12:00       ` Ramkumar Ramachandra
2013-07-08 13:13         ` Duy Nguyen
2013-07-08 13:37           ` Ramkumar Ramachandra
2013-07-09  2:55             ` Duy Nguyen
2013-07-09 10:32               ` Ramkumar Ramachandra
2013-07-10 11:16             ` Jeff King
2013-07-08 16:40           ` Junio C Hamano
2013-07-10 11:04     ` Jeff King
2013-07-11 16:35       ` Junio C Hamano
2013-07-07 21:15   ` brian m. carlson
2013-07-10 10:57     ` Jeff King
2013-07-07 10:14 ` [PATCH 4/4] pack-revindex: radix-sort the revindex Jeff King
2013-07-07 23:52   ` Shawn Pearce
2013-07-08  7:57     ` Jeff King
2013-07-08 15:38       ` Shawn Pearce
2013-07-08 20:50   ` Brandon Casey [this message]
2013-07-08 21:35     ` Brandon Casey
2013-07-10 10:57       ` Jeff King
2013-07-10 10:52     ` Jeff King
2013-07-10 11:34 ` [PATCHv2 00/10] cat-file formats/on-disk sizes Jeff King
2013-07-10 11:35   ` [PATCH 01/10] zero-initialize object_info structs Jeff King
2013-07-10 11:35   ` [PATCH 02/10] teach sha1_object_info_extended a "disk_size" query Jeff King
2013-07-10 11:36   ` [PATCH 03/10] t1006: modernize output comparisons Jeff King
2013-07-10 11:38   ` [PATCH 04/10] cat-file: teach --batch to stream blob objects Jeff King
2013-07-10 11:38   ` [PATCH 05/10] cat-file: refactor --batch option parsing Jeff King
2013-07-10 11:45   ` [PATCH 06/10] cat-file: add --batch-check=<format> Jeff King
2013-07-10 11:57     ` Eric Sunshine
2013-07-10 14:51     ` Ramkumar Ramachandra
2013-07-11 11:24       ` Jeff King
2013-07-10 11:46   ` [PATCH 07/10] cat-file: add %(objectsize:disk) format atom Jeff King
2013-07-10 11:48   ` [PATCH 08/10] cat-file: split --batch input lines on whitespace Jeff King
2013-07-10 15:29     ` Ramkumar Ramachandra
2013-07-11 11:36       ` Jeff King
2013-07-11 17:42         ` Junio C Hamano
2013-07-11 20:45         ` [PATCHv3 " Jeff King
2013-07-10 11:50   ` [PATCH 09/10] pack-revindex: use unsigned to store number of objects Jeff King
2013-07-10 11:55   ` [PATCH 10/10] pack-revindex: radix-sort the revindex Jeff King
2013-07-10 12:00     ` Jeff King
2013-07-10 13:17     ` Ramkumar Ramachandra
2013-07-11 11:03       ` Jeff King
2013-07-10 17:10     ` Brandon Casey
2013-07-11 11:17       ` Jeff King
2013-07-11 12:16     ` [PATCHv3 " Jeff King
2013-07-11 21:12       ` Brandon Casey

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

  List information: http://vger.kernel.org/majordomo-info.html

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CA+sFfMfQcnx+OGNd+v7NJC5zSXg2OR1QiLyRSqDjXD0zb4mvtA@mail.gmail.com \
    --to=drafnel@gmail.com \
    --cc=git@vger.kernel.org \
    --cc=peff@peff.net \
    --subject='Re: [PATCH 4/4] pack-revindex: radix-sort the revindex' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

Code repositories for project(s) associated with this inbox:

	https://80x24.org/mirrors/git.git

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).