bug#27560: 26.0.50; seq-uniq is slow

bug#27560: 26.0.50; seq-uniq is slow

[Nicolas Petton]

[-- Attachment #1: Type: text/plain, Size: 733 bytes --]


Oleh (in Cc) did some benchmarks in May 2015, and `seq-uniq' is quite
slow:

    (setq cands (locate-file-completion-table
		 load-path (get-load-suffixes) "" nil t))
    (length cands)
    5357
    (length (cl-remove-duplicates cands :test 'equal))
    2481
    (benchmark-run (cl-remove-duplicates cands :test 'equal))
    (0.67873101 0 0.0)
    (benchmark-run (helm-fast-remove-dups cands :test 'equal))
    (0.001350054 0 0.0)
    (benchmark-run (seq-uniq cands 'equal))
    (5.270219822 27 2.396615401000002)

One easy way to make it faster would be to use `cl-remove-duplicates'
for sequences (sequencep), and default to the current implementation for
other seqp data structures (which means stream.el currently AFAIK).

Nico

[-- Attachment #2.1: Type: text/plain, Size: 0 bytes --]



[-- Attachment #2.2: signature.asc --]
[-- Type: application/pgp-signature, Size: 472 bytes --]

bug#27560: 26.0.50; seq-uniq is slow

[Tino Calancha]

Nicolas Petton <nicolas@petton.fr> writes:

> Oleh (in Cc) did some benchmarks in May 2015, and `seq-uniq' is quite
> slow:
>
>     (setq cands (locate-file-completion-table
> 		 load-path (get-load-suffixes) "" nil t))
>     (length cands)
>     5357
>     (length (cl-remove-duplicates cands :test 'equal))
>     2481
>     (benchmark-run (cl-remove-duplicates cands :test 'equal))
>     (0.67873101 0 0.0)
>     (benchmark-run (helm-fast-remove-dups cands :test 'equal))
>     (0.001350054 0 0.0)
>     (benchmark-run (seq-uniq cands 'equal))
>     (5.270219822 27 2.396615401000002)
>
> One easy way to make it faster would be to use `cl-remove-duplicates'
> for sequences (sequencep), and default to the current implementation for
> other seqp data structures (which means stream.el currently AFAIK).
The numbers above are Oleh's numbers from 2015.
If i run similar benchmarks now i don't get such impressive
difference between cl-lib/seq:

(setq cands (locate-file-completion-table
		 load-path (get-load-suffixes) "" nil t))
("cc-compat" "vi" "longlines" "rcompile" "eudcb-ph" "yow" "fast-lock" "cl-compat" "mouse-sel" "tpu-mapper" "otodo-mode" "levents" ...)
(length cands)
5317
(length (cl-remove-duplicates cands :test 'equal))
1716

;; Each benchmark run 5 times:

(benchmark-run 1 (helm-fast-remove-dups cands :test 'equal))
(0.003539776 0 0.0)
(0.007002079 0 0.0)
(0.006296864 0 0.0)
(0.002854921 0 0.0)
(0.009388854 0 0.0)

(benchmark-run 1 (cl-remove-duplicates cands :test 'equal))
(0.558590947 0 0.0)
(0.554376583 0 0.0)
(0.554984501 0 0.0)
(0.583015589 0 0.0)
(0.565781774 0 0.0)

(benchmark-run 1 (seq-uniq cands 'equal))
(0.36366049 0 0.0)
(0.37484882 0 0.0)
(0.604267587 1 0.22640233999999282)
(0.360611939 0 0.0)
(0.392288023 0 0.0)

Tino

bug#27560: 26.0.50; seq-uniq is slow

[Nicolas Petton]

[-- Attachment #1: Type: text/plain, Size: 423 bytes --]

Tino Calancha <tino.calancha@gmail.com> writes:

> The numbers above are Oleh's numbers from 2015.

Yes, but AFAIR `seq-uniq' hasn't changed since then.

> If i run similar benchmarks now i don't get such impressive
> difference between cl-lib/seq:

Indeed, I will run some benchmarks as well. In your benchmarks, cl-lib
is still much faster than seq-uniq, even if it's not as impressive as
Oleh's benchmark.

Cheers,
Nico

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 472 bytes --]

bug#27560: 26.0.50; seq-uniq is slow

[Nicolas Petton]

[-- Attachment #1: Type: text/plain, Size: 262 bytes --]

Nicolas Petton <nicolas@petton.fr> writes:

> Indeed, I will run some benchmarks as well. In your benchmarks, cl-lib
> is still much faster than seq-uniq, even if it's not as impressive as
> Oleh's benchmark.

It looks like I read the benchmark results wrong :)

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 472 bytes --]

bug#27560: 26.0.50; seq-uniq is slow

[Nicolas Petton]

[-- Attachment #1: Type: text/plain, Size: 445 bytes --]

Nicolas Petton <nicolas@petton.fr> writes:

>> Indeed, I will run some benchmarks as well. In your benchmarks, cl-lib
>> is still much faster than seq-uniq, even if it's not as impressive as
>> Oleh's benchmark.
>
> It looks like I read the benchmark results wrong :)

seq-uniq seems to be much faster than in 2015, while its definition
basically hasn't changed.  I won't investigate much further as I'm happy
with then benchmark results.

Nico

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 472 bytes --]

bug#27560: 26.0.50; seq-uniq is slow

[Tino Calancha]

On Mon, 3 Jul 2017, Nicolas Petton wrote:

> Nicolas Petton <nicolas@petton.fr> writes:
>
>>> Indeed, I will run some benchmarks as well. In your benchmarks, cl-lib
>>> is still much faster than seq-uniq, even if it's not as impressive as
>>> Oleh's benchmark.
>>
>> It looks like I read the benchmark results wrong :)
>
> seq-uniq seems to be much faster than in 2015, while its definition
> basically hasn't changed.  I won't investigate much further as I'm happy
> with then benchmark results.
I am quite amazed with the 27 garbage collections that Oleh found:
>     (benchmark-run (seq-uniq cands 'equal))
>     (5.270219822 27 2.396615401000002)

I might get occasionally 1 gc, tipically 0: 27 demands a double checking.
Tino