From: Jeff King <peff@peff.net>
To: Taylor Blau <me@ttaylorr.com>
Cc: git@vger.kernel.org, Patrick Steinhardt <ps@pks.im>,
Junio C Hamano <gitster@pobox.com>
Subject: Re: [PATCH 00/24] pack-objects: multi-pack verbatim reuse
Date: Thu, 21 Dec 2023 06:13:33 -0500 [thread overview]
Message-ID: <20231221111333.GE570888@coredump.intra.peff.net> (raw)
In-Reply-To: <ZXxy1USjjjAbBi++@nand.local>
On Fri, Dec 15, 2023 at 10:37:57AM -0500, Taylor Blau wrote:
> On Tue, Dec 12, 2023 at 03:12:38AM -0500, Jeff King wrote:
> > So my question is: how much of what you're seeing is from (1) and (2),
> > and how much is from (3)? Because there are other ways to trigger (3),
> > such as lowering the window size. For example, if you try your same
> > packing example with --window=0, how do the CPU and output size compare
> > to the results of your series? (I'd also check peak memory usage).
>
> Interesting question! Here are some preliminary numbers on my machine
> (which runs Debian unstable with a Intel Xenon W-2255 CPU @ 3.70GHz and
> 64GB of RAM).
>
> I ran the following hyperfine command on my testing repository, which
> has the Git repository broken up into ~75 packs or so:
Thanks for running these tests. The results are similar to what
expected, which is: yes, most of your CPU savings are from skipping
deltas, but not all.
Here's what I see (which I think is mostly redundant with what you've
said, but I just want to lay out my line of thinking). I'll reorder your
quoted sections a bit as I go:
> Benchmark 2: multi-pack reuse, pack.window=0
> [...]
> Time (mean ± σ): 1.075 s ± 0.005 s [User: 0.990 s, System: 0.188 s]
> Range (min … max): 1.071 s … 1.088 s 10 runs
>
> Benchmark 4: multi-pack reuse, pack.window=10
> [...]
> Time (mean ± σ): 1.028 s ± 0.002 s [User: 1.150 s, System: 0.184 s]
> Range (min … max): 1.026 s … 1.032 s 10 runs
OK, so when we're doing more full ("multi") reuse, the pack window
doesn't make a big difference either way. You didn't show the stderr
from each, but presumably most of the objects are hitting the "reuse"
path, and only a few are deltas (and that is backed up by the fact that
doing deltas only gives us a slight improvement in the output size:
> Benchmark 2: multi-pack reuse, pack.window=0
> 268.670 MB
> Benchmark 4: multi-pack reuse, pack.window=10
> 266.473 MB
Comparing the runs with less reuse:
> Benchmark 1: single-pack reuse, pack.window=0
> [...]
> Time (mean ± σ): 1.248 s ± 0.004 s [User: 1.160 s, System: 0.188 s]
> Range (min … max): 1.244 s … 1.259 s 10 runs
>
> Benchmark 3: single-pack reuse, pack.window=10
> [...]
> Time (mean ± σ): 6.281 s ± 0.024 s [User: 43.727 s, System: 0.492 s]
> Range (min … max): 6.252 s … 6.326 s 10 runs
there obviously is a huge amount of time saved by not doing deltas, but
we pay for it with a much bigger pack:
> Benchmark 1: single-pack reuse, pack.window=0
> 264.443 MB
> Benchmark 3: single-pack reuse, pack.window=10
> 194.355 MB
But of course that "much bigger" pack is about the same size as the one
we get from doing multi-pack reuse. Which is not surprising, because
both are avoiding looking for new deltas (and the packs after the
preferred one probably have mediocre deltas).
So I do actually think that disabling pack.window gives you a
similar-ish tradeoff to expanding the pack-reuse code (~6s down to ~1s,
and a 36% embiggening of the resulting pack size).
Which implies that one option is to scrap your entire series and just
set pack.window. Basically comparing multi/10 (your patches) to single/0
(a hypothetical config option), which have similar run-times and pack
sizes.
But that's not quite the whole story. There is still a CPU improvement
in your series (1.2s vs 1.0s, a 20% speedup). And as I'd expect, a
memory improvement from avoiding the extra book-keeping (almost 10%):
> Benchmark 1: single-pack reuse, pack.window=0
> 354.224 MB (max RSS)
> Benchmark 4: multi-pack reuse, pack.window=10
> 328.786 MB (max RSS)
So while it's a lot less code to just set the window size, I do think
those improvements are worth it. And really, it's the same tradeoff we
make for the single-pack case (i.e., one could argue that we
could/should rip out the verbatim-reuse code entirely in favor of just
tweaking the window size).
> It's pretty close between multi-pack reuse with a window size of 0 and
> a window size of 10. If you want to optimize for pack size, you could
> trade a ~4% reduction in pack size for a ~1% increase in peak memory
> usage.
I think if you want to optimize for pack size, you should consider
repacking all-into-one to get better on-disk deltas. ;) I know that's
easier said than done when the I/O costs are significant. I do wonder if
storing thin packs on disk would let us more cheaply reach a state that
could serve optimal-ish packs without spending CPU computing bespoke
deltas for each client. But that's a much larger topic.
-Peff
next prev parent reply other threads:[~2023-12-21 11:13 UTC|newest]
Thread overview: 107+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-11-28 19:07 [PATCH 00/24] pack-objects: multi-pack verbatim reuse Taylor Blau
2023-11-28 19:07 ` [PATCH 01/24] pack-objects: free packing_data in more places Taylor Blau
2023-11-30 10:18 ` Patrick Steinhardt
2023-11-30 19:08 ` Taylor Blau
2023-11-28 19:07 ` [PATCH 02/24] pack-bitmap-write: deep-clear the `bb_commit` slab Taylor Blau
2023-11-30 10:18 ` Patrick Steinhardt
2023-11-30 19:11 ` Taylor Blau
2023-12-12 7:04 ` Jeff King
2023-12-12 16:48 ` Taylor Blau
2023-11-28 19:08 ` [PATCH 03/24] pack-bitmap: plug leak in find_objects() Taylor Blau
2023-12-12 7:04 ` Jeff King
2023-11-28 19:08 ` [PATCH 04/24] midx: factor out `fill_pack_info()` Taylor Blau
2023-11-30 10:18 ` Patrick Steinhardt
2023-11-30 19:19 ` Taylor Blau
2023-11-28 19:08 ` [PATCH 05/24] midx: implement `DISP` chunk Taylor Blau
2023-11-30 10:18 ` Patrick Steinhardt
2023-11-30 19:27 ` Taylor Blau
2023-12-03 13:15 ` Junio C Hamano
2023-12-05 19:26 ` Taylor Blau
2023-12-09 1:40 ` Junio C Hamano
2023-12-09 2:30 ` Taylor Blau
2023-12-12 8:03 ` Jeff King
2023-12-13 18:28 ` Taylor Blau
2023-12-13 19:20 ` Junio C Hamano
2023-11-28 19:08 ` [PATCH 06/24] midx: implement `midx_locate_pack()` Taylor Blau
2023-11-28 19:08 ` [PATCH 07/24] midx: implement `--retain-disjoint` mode Taylor Blau
2023-11-30 10:18 ` Patrick Steinhardt
2023-11-30 19:29 ` Taylor Blau
2023-12-01 8:02 ` Patrick Steinhardt
2023-11-28 19:08 ` [PATCH 08/24] pack-objects: implement `--ignore-disjoint` mode Taylor Blau
2023-11-30 10:18 ` Patrick Steinhardt
2023-11-30 19:32 ` Taylor Blau
2023-12-01 8:17 ` Patrick Steinhardt
2023-12-01 19:58 ` Taylor Blau
2023-11-28 19:08 ` [PATCH 09/24] repack: implement `--extend-disjoint` mode Taylor Blau
2023-12-07 13:13 ` Patrick Steinhardt
2023-12-07 20:28 ` Taylor Blau
2023-12-08 8:19 ` Patrick Steinhardt
2023-12-08 22:48 ` Taylor Blau
2023-12-11 8:18 ` Patrick Steinhardt
2023-12-11 19:59 ` Taylor Blau
2023-11-28 19:08 ` [PATCH 10/24] pack-bitmap: pass `bitmapped_pack` struct to pack-reuse functions Taylor Blau
2023-12-07 13:13 ` Patrick Steinhardt
2023-12-07 20:34 ` Taylor Blau
2023-12-08 8:19 ` Patrick Steinhardt
2023-11-28 19:08 ` [PATCH 11/24] pack-bitmap: simplify `reuse_partial_packfile_from_bitmap()` signature Taylor Blau
2023-12-07 13:13 ` Patrick Steinhardt
2023-12-07 14:36 ` Taylor Blau
2023-11-28 19:08 ` [PATCH 12/24] pack-bitmap: return multiple packs via `reuse_partial_packfile_from_bitmap()` Taylor Blau
2023-11-28 19:08 ` [PATCH 13/24] pack-objects: parameterize pack-reuse routines over a single pack Taylor Blau
2023-11-28 19:08 ` [PATCH 14/24] pack-objects: keep track of `pack_start` for each reuse pack Taylor Blau
2023-12-07 13:13 ` Patrick Steinhardt
2023-12-07 20:43 ` Taylor Blau
2023-11-28 19:08 ` [PATCH 15/24] pack-objects: pass `bitmapped_pack`'s to pack-reuse functions Taylor Blau
2023-11-28 19:08 ` [PATCH 16/24] pack-objects: prepare `write_reused_pack()` for multi-pack reuse Taylor Blau
2023-12-07 13:13 ` Patrick Steinhardt
2023-12-07 20:47 ` Taylor Blau
2023-11-28 19:08 ` [PATCH 17/24] pack-objects: prepare `write_reused_pack_verbatim()` " Taylor Blau
2023-11-28 19:08 ` [PATCH 18/24] pack-objects: include number of packs reused in output Taylor Blau
2023-11-28 19:08 ` [PATCH 19/24] pack-bitmap: prepare to mark objects from multiple packs for reuse Taylor Blau
2023-11-28 19:08 ` [PATCH 20/24] pack-objects: add tracing for various packfile metrics Taylor Blau
2023-11-28 19:08 ` [PATCH 21/24] t/test-lib-functions.sh: implement `test_trace2_data` helper Taylor Blau
2023-11-28 19:08 ` [PATCH 22/24] pack-objects: allow setting `pack.allowPackReuse` to "single" Taylor Blau
2023-11-28 19:08 ` [PATCH 23/24] pack-bitmap: reuse objects from all disjoint packs Taylor Blau
2023-11-28 19:08 ` [PATCH 24/24] t/perf: add performance tests for multi-pack reuse Taylor Blau
2023-11-30 10:18 ` [PATCH 00/24] pack-objects: multi-pack verbatim reuse Patrick Steinhardt
2023-11-30 19:39 ` Taylor Blau
2023-12-01 8:31 ` Patrick Steinhardt
2023-12-01 20:02 ` Taylor Blau
2023-12-04 8:49 ` Patrick Steinhardt
2023-12-12 8:12 ` Jeff King
2023-12-15 15:37 ` Taylor Blau
2023-12-21 11:13 ` Jeff King [this message]
2024-01-04 22:22 ` Taylor Blau
2023-12-14 22:23 ` [PATCH v2 00/26] " Taylor Blau
2023-12-14 22:23 ` [PATCH v2 01/26] pack-objects: free packing_data in more places Taylor Blau
2023-12-14 22:23 ` [PATCH v2 02/26] pack-bitmap-write: deep-clear the `bb_commit` slab Taylor Blau
2023-12-14 22:23 ` [PATCH v2 03/26] pack-bitmap: plug leak in find_objects() Taylor Blau
2023-12-14 22:23 ` [PATCH v2 04/26] midx: factor out `fill_pack_info()` Taylor Blau
2023-12-14 22:23 ` [PATCH v2 05/26] midx: implement `BTMP` chunk Taylor Blau
2023-12-14 22:23 ` [PATCH v2 06/26] midx: implement `midx_locate_pack()` Taylor Blau
2023-12-14 22:23 ` [PATCH v2 07/26] pack-bitmap: pass `bitmapped_pack` struct to pack-reuse functions Taylor Blau
2023-12-14 22:23 ` [PATCH v2 08/26] ewah: implement `bitmap_is_empty()` Taylor Blau
2023-12-14 22:24 ` [PATCH v2 09/26] pack-bitmap: simplify `reuse_partial_packfile_from_bitmap()` signature Taylor Blau
2023-12-14 22:24 ` [PATCH v2 10/26] pack-bitmap: return multiple packs via `reuse_partial_packfile_from_bitmap()` Taylor Blau
2023-12-14 22:24 ` [PATCH v2 11/26] pack-objects: parameterize pack-reuse routines over a single pack Taylor Blau
2023-12-14 22:24 ` [PATCH v2 12/26] pack-objects: keep track of `pack_start` for each reuse pack Taylor Blau
2023-12-14 22:24 ` [PATCH v2 13/26] pack-objects: pass `bitmapped_pack`'s to pack-reuse functions Taylor Blau
2023-12-14 22:24 ` [PATCH v2 14/26] pack-objects: prepare `write_reused_pack()` for multi-pack reuse Taylor Blau
2023-12-14 22:24 ` [PATCH v2 15/26] pack-objects: prepare `write_reused_pack_verbatim()` " Taylor Blau
2023-12-14 22:24 ` [PATCH v2 16/26] pack-objects: include number of packs reused in output Taylor Blau
2023-12-14 22:24 ` [PATCH v2 17/26] git-compat-util.h: implement checked size_t to uint32_t conversion Taylor Blau
2023-12-14 22:24 ` [PATCH v2 18/26] midx: implement `midx_preferred_pack()` Taylor Blau
2023-12-14 22:24 ` [PATCH v2 19/26] pack-revindex: factor out `midx_key_to_pack_pos()` helper Taylor Blau
2023-12-14 22:24 ` [PATCH v2 20/26] pack-revindex: implement `midx_pair_to_pack_pos()` Taylor Blau
2023-12-14 22:24 ` [PATCH v2 21/26] pack-bitmap: prepare to mark objects from multiple packs for reuse Taylor Blau
2023-12-14 22:24 ` [PATCH v2 22/26] pack-objects: add tracing for various packfile metrics Taylor Blau
2023-12-14 22:24 ` [PATCH v2 23/26] t/test-lib-functions.sh: implement `test_trace2_data` helper Taylor Blau
2023-12-14 22:24 ` [PATCH v2 24/26] pack-objects: allow setting `pack.allowPackReuse` to "single" Taylor Blau
2023-12-14 22:24 ` [PATCH v2 25/26] pack-bitmap: enable reuse from all bitmapped packs Taylor Blau
2023-12-14 22:24 ` [PATCH v2 26/26] t/perf: add performance tests for multi-pack reuse Taylor Blau
2023-12-15 0:06 ` [PATCH v2 00/26] pack-objects: multi-pack verbatim reuse Junio C Hamano
2023-12-15 0:38 ` Taylor Blau
2023-12-15 0:40 ` Junio C Hamano
2023-12-15 1:25 ` Taylor Blau
2023-12-21 10:51 ` Jeff King
2024-01-04 22:24 ` Taylor Blau
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
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20231221111333.GE570888@coredump.intra.peff.net \
--to=peff@peff.net \
--cc=git@vger.kernel.org \
--cc=gitster@pobox.com \
--cc=me@ttaylorr.com \
--cc=ps@pks.im \
/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
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
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 NNTP newsgroup(s).