The previous patch almost halved the number of heap allocations for the
sort subcommand. Reduce it further by using a mem_pool for the line
objects.
Note that t/perf/run can't be used directly to compare two versions of
test-mergesort because it always runs the helpers from the checked-out
version. So I hand-merged the results of separate runs before and with
this patch:
macOS 12.5.1 on M1:
0071.12: DEFINE_LIST_SORT unsorted 0.22(0.20+0.01) 0.21(0.19+0.01)
0071.14: DEFINE_LIST_SORT sorted 0.10(0.08+0.01) 0.10(0.08+0.01)
0071.16: DEFINE_LIST_SORT reversed 0.10(0.08+0.01) 0.10(0.08+0.01)
Git SDK 64-bit on Windows 11 21H2 on Ryzen 7 5800H:
0071.12: DEFINE_LIST_SORT unsorted 0.54(0.00+0.06) 0.44(0.01+0.06)
0071.14: DEFINE_LIST_SORT sorted 0.21(0.03+0.03) 0.19(0.04+0.01)
0071.16: DEFINE_LIST_SORT reversed 0.21(0.01+0.04) 0.19(0.04+0.04)
Debian bullseye on WSL2 on the same system:
0071.12: DEFINE_LIST_SORT unsorted 0.29(0.27+0.01) 0.22(0.19+0.02)
0071.14: DEFINE_LIST_SORT sorted 0.07(0.06+0.01) 0.06(0.04+0.02)
0071.16: DEFINE_LIST_SORT reversed 0.07(0.04+0.03) 0.06(0.04+0.02)
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The sort subcommand of test-mergesort is used to test the performance of
sorting linked lists. It reads lines from stdin, sorts them and prints
the result to stdout. Two heap allocations are done per line: One for
the linked list item and one for the actual line string. That imposes a
significant amount of allocation overhead.
Reduce it by doing the same as the sort subcommand of test-string-list,
namely to read the whole input file into a single buffer and then split
it in-place.
Note that t/perf/run can't be used directly to compare two versions of
test-mergesort because it always runs the helpers from the checked-out
version. So I hand-merged the results of separate runs before and with
this patch:
macOS 12.5.1 on M1:
0071.12: DEFINE_LIST_SORT unsorted 0.23(0.20+0.01) 0.22(0.20+0.01)
0071.14: DEFINE_LIST_SORT sorted 0.12(0.10+0.01) 0.10(0.08+0.01)
0071.16: DEFINE_LIST_SORT reversed 0.12(0.10+0.01) 0.10(0.08+0.01)
Git SDK 64-bit on Windows 11 21H2 on Ryzen 7 5800H:
0071.12: DEFINE_LIST_SORT unsorted 0.71(0.00+0.03) 0.54(0.00+0.06)
0071.14: DEFINE_LIST_SORT sorted 0.42(0.00+0.04) 0.21(0.03+0.03)
0071.16: DEFINE_LIST_SORT reversed 0.42(0.06+0.01) 0.21(0.01+0.04)
Debian bullseye on WSL2 on the same system:
0071.12: DEFINE_LIST_SORT unsorted 0.41(0.39+0.02) 0.29(0.27+0.01)
0071.14: DEFINE_LIST_SORT sorted 0.11(0.08+0.02) 0.07(0.06+0.01)
0071.16: DEFINE_LIST_SORT reversed 0.11(0.08+0.02) 0.07(0.04+0.03)
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Build a typed sort function for the mergesort performance test tool
using DEFINE_LIST_SORT instead of calling llist_mergesort(). This gets
rid of the next pointer accessor functions and improves the performance
at the cost of a slightly higher object text size.
Before:
0071.12: llist_mergesort() unsorted 0.24(0.22+0.01)
0071.14: llist_mergesort() sorted 0.12(0.10+0.01)
0071.16: llist_mergesort() reversed 0.12(0.10+0.01)
__TEXT __DATA __OBJC others dec hex
6407 276 0 24701 31384 7a98 t/helper/test-mergesort.o
With this patch:
0071.12: DEFINE_LIST_SORT unsorted 0.22(0.21+0.01)
0071.14: DEFINE_LIST_SORT sorted 0.11(0.10+0.01)
0071.16: DEFINE_LIST_SORT reversed 0.11(0.10+0.01)
__TEXT __DATA __OBJC others dec hex
6615 276 0 25832 32723 7fd3 t/helper/test-mergesort.o
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Define a typed sort function using DEFINE_LIST_SORT_DEBUG for the
mergesort sanity check instead of using llist_mergesort(). This gets
rid of the next pointer accessor functions and improves the performance
at the cost of slightly bigger object text.
Before:
Benchmark 1: t/helper/test-tool mergesort test
Time (mean ± σ): 108.4 ms ± 0.2 ms [User: 106.7 ms, System: 1.2 ms]
Range (min … max): 108.0 ms … 108.8 ms 27 runs
__TEXT __DATA __OBJC others dec hex
6251 276 0 23172 29699 7403 t/helper/test-mergesort.o
With this patch:
Benchmark 1: t/helper/test-tool mergesort test
Time (mean ± σ): 94.0 ms ± 0.2 ms [User: 92.4 ms, System: 1.1 ms]
Range (min … max): 93.7 ms … 94.5 ms 31 runs
__TEXT __DATA __OBJC others dec hex
6407 276 0 24701 31384 7a98 t/helper/test-mergesort.o
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use MINSTD to generate pseudo-random numbers consistently instead of
using rand(3), whose output can vary from system to system, and reset
its seed before filling in the test values. This gives repeatable
results across versions and systems, which simplifies sharing and
comparing of results between developers.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a mode that turns a sorted list into adversarial input for a
bottom-up mergesort implementation that doubles the length of sorted
sublists at each level -- like our llist_mergesort().
While unriffle mode splits the list in half at each recursion step,
unriffle_skewed splits it into 2^l items and the rest, with 2^l being
the highest power of two smaller than the number of items and thus
2^l >= rest. The rest is unriffled with the tail of the first half to
require a merge to compare the maximum number of elements.
It complements the unriffle mode, which targets balanced merges. If
the number of elements is a power of two then both actually produce the
same result, as 2^l == rest == n/2 at each recursion step in that case.
Here are the results:
$ t/helper/test-tool mergesort test | awk '
$7 > max[$3] {max[$3] = $7; line[$3] = $0}
END {for (n in line) print line[n]}
'
distribut mode n m get_next set_next compare verdict
sawtooth unriffle_skewed 100 128 1184 700 589 OK
sawtooth unriffle_skewed 1023 1024 16373 10230 9207 OK
sawtooth unriffle 1024 1024 16384 10240 9217 OK
sawtooth unriffle_skewed 1025 2048 18454 11275 10241 OK
The sawtooth distribution with m>=n produces a sorted list and
unriffle_skewed mode turns it into adversarial input for unbalanced
merges, which it wins in all cases except for n=1024 -- the resulting
list is the same, but unriffle is tested before unriffle_skewed, so its
result is selected by the AWK script.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a mode that turns sorted items into adversarial input for mergesort.
Do that by running mergesort in reverse and rearranging the items in
such a way that each merge needs the maximum number of operations to
undo it.
To riffle is a card shuffling technique and involves splitting a deck
into two and then to interleave them. A perfect riffle takes one card
from each half in turn. That's similar to the most expensive merge,
which has to take one item from each sublist in turn, which requires the
maximum number of comparisons (n-1).
So unriffle does that in reverse, i.e. it generates the first sublist
out of the items at even indexes and the second sublist out of the items
at odd indexes, without changing their order in any other way. Done
recursively until we reach the trivial sublist length of one, this
twists the list into an order that requires the maximum effort for
mergesort to untangle.
As a baseline, here are the rand distributions with the highest number
of comparisons from "test-tool mergesort test":
$ t/helper/test-tool mergesort test | awk '
NR > 1 && $1 != "rand" {next}
$7 > max[$3] {max[$3] = $7; line[$3] = $0}
END {for (n in line) print line[n]}
'
distribut mode n m get_next set_next compare verdict
rand copy 100 32 1184 700 569 OK
rand reverse_1st_half 1023 256 16373 10230 8976 OK
rand reverse_1st_half 1024 512 16384 10240 8993 OK
rand dither 1025 64 18454 11275 9970 OK
And here are the most expensive ones overall:
$ t/helper/test-tool mergesort test | awk '
$7 > max[$3] {max[$3] = $7; line[$3] = $0}
END {for (n in line) print line[n]}
'
distribut mode n m get_next set_next compare verdict
stagger reverse 100 64 1184 700 580 OK
sawtooth unriffle 1023 1024 16373 10230 9179 OK
sawtooth unriffle 1024 1024 16384 10240 9217 OK
stagger unriffle 1025 2048 18454 11275 10241 OK
The sawtooth distribution with m>=n generates a sorted list. The
unriffle mode is designed to turn that into adversarial input for
mergesort, and that checks out for n=1023 and n=1024, where it produces
the list that requires the most comparisons.
Item counts that are not powers of two have other winners, and that's
because unriffle recursively splits lists into equal-sized halves, while
llist_mergesort() splits them into the biggest power of two smaller than
n and the rest, e.g. for n=1025 it sorts the first 1024 separately and
finally merges them to the last item.
So unriffle mode works as designed for the intended use case, but to
consistently generate adversarial input for unbalanced merges we need
something else.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a subcommand for printing test data. It can be used to generate
special test cases and feed them into the sort subcommand or sort(1) for
performance measurements. It may also be useful to illustrate the
effect of distributions, modes and their parameters.
It generates n integers with the specified distribution and its
distribution-specific parameter m. E.g. m is the maximum value for
the plateau distribution and the length and height of individual teeth
of the sawtooth distribution.
The generated values are printed as zero-padded eight-digit hexadecimal
numbers to make sure alphabetic and numeric order are the same.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Adapt the qsort certification program from "Engineering a Sort Function"
by Bentley and McIlroy for testing our linked list sort function. It
generates several lists with various distribution patterns and counts
the number of operations llist_mergesort() needs to order them. It
compares the result to the output of a trusted sort function (qsort(1))
and also checks if the sort is stable.
Also add a test script that makes use of the new subcommand.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Give the code for sorting a text file its own sub-command. This allows
extending the helper, which we'll do in the following patches.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Strip line ending characters to make sure empty lines are sorted like
sort(1) does.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This keeps top dir a bit less crowded. And because these programs are
for testing purposes, it makes sense that they stay somewhere in t/
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Elijah Newren
René Scharfe
Nguyễn Thái Ngọc Duy
Junio C Hamano