Word Count program/benchmark performance

[Pawel Mosakowski]

Hello,

I am a Racket beginner and I have come across this article:

https://benhoyt.com/writings/count-words/

This is my attempt at solving the challenge:

https://pastebin.com/kL16w5Hc

However when I have benchmarked it, it takes ~21 seconds to run compared to the Python and Ruby versions which take around 3-4 seconds.

I understand that both Ruby and Python probably have the string operations and hash tables implemented in optimized C but is there anything I can do to improve performance of my program?

Many thanks for all help and suggestions.

[David Storrs]

Hi Pawel,

I'll take a look at the code later, but did that 21 seconds include startup time for the interpreter?

--
You received this message because you are subscribed to the Google Groups "Racket Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to racket-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/118c1340-66d1-421d-92a4-6b66c56cb88fn%40googlegroups.com.

[Pawel Mosakowski]

Hi David,

Yes, the 21 seconds includes the interpreter startup time. I have done a simple test to see how long it takes:

$ time racket -e '(displayln "Hello, world")'
Hello, world

real    0m0.479s
user    0m0.449s

sys    0m0.030s

I have also put my code inside a main function and profiled it:

Profiling results
-----------------
  Total cpu time observed: 20910ms (out of 20970ms)
  Number of samples taken: 382 (once every 55ms)
  (Hiding functions with self<1.0% and local<2.0%: 1 of 12 hidden)

==============================================================
                                  Caller
 Idx    Total         Self      Name+src                Local%
        ms(pct)       ms(pct)     Callee
==============================================================
 [1] 20910(100.0%)     0(0.0%)  [running body] ...word-occurences-profile.rkt":##f
                                  profile-thunk [2]     100.0%
--------------------------------------------------------------
                                  [running body] [1]    100.0%
 [2] 20910(100.0%)     0(0.0%)  profile-thunk ...ket/pkgs/profile-lib/main.rkt:9:0
                                  run [3]               100.0%
--------------------------------------------------------------
                                  profile-thunk [2]     100.0%
 [3] 20910(100.0%)     0(0.0%)  run ...share/racket/pkgs/profile-lib/main.rkt:39:2
                                  main [4]              100.0%
--------------------------------------------------------------
                                  run [3]               100.0%
 [4] 20910(100.0%)    50(0.2%)  main ...cket/count-word-occurences-profile.rkt:5:0
                                  read-from-stdin-it [5] 98.5%
                                  ??? [6]                 0.2%
--------------------------------------------------------------
                                  main [4]              100.0%
 [5] 20606(98.5%)  11796(56.4%) read-from-stdin-it ...-occurences-profile.rkt:19:6
                                  internal-split [7]     42.8%
--------------------------------------------------------------
                                  main [4]              100.0%
 [6]    51(0.2%)       0(0.0%)  ??? ...cket/collects/racket/private/sort.rkt:369:3
                                  generic-sort/key [8]  100.0%
--------------------------------------------------------------
                                  read-from-stdin-it [5]100.0%
 [7]  8810(42.1%)   3528(16.9%) internal-split ...collects/racket/string.rkt:117:0
                                  regexp-split [9]       59.9%
--------------------------------------------------------------
                                  ??? [6]               100.0%
 [8]    51(0.2%)       0(0.0%)  generic-sort/key .../racket/private/sort.rkt:156:2
                                  copying-mergesort [10]100.0%
--------------------------------------------------------------
                                  internal-split [7]    100.0%
 [9]  5282(25.3%)   2810(13.4%) regexp-split ...ts/racket/private/string.rkt:338:2
                                  loop [11]              46.8%
--------------------------------------------------------------
                                  generic-sort/key [8]   10.0%
                                  copying-mergesort [10] 90.0%
[10]    51(0.2%)      51(0.2%)  copying-mergesort ...racket/private/sort.rkt:129:8
                                  copying-mergesort [10] 90.0%
--------------------------------------------------------------
                                  regexp-split [9]      100.0%
[11]  2471(11.8%)   2471(11.8%) loop ...t/collects/racket/private/string.rkt:169:7
--------------------------------------------------------------

Kind regards,

Pawel

[Sam Tobin-Hochstadt]

Here's a somewhat-optimized version of the code:

#lang racket/base
(require racket/string racket/vector racket/port)

(define h (make-hash))

(time
(for* ([l (in-lines)]
[w (in-list (string-split l))]
[w* (in-value (string-downcase w))])
(hash-update! h w* add1 0)))

(define v
(time
(for/vector #:length (hash-count h)
([(k v) (in-hash h)])
(cons k v))))
(time (vector-sort! v > #:key cdr))
(define p (current-output-port) #;(open-output-nowhere))
(time
(for ([pair (in-vector v)])
(write-string (car pair) p)
(write-string (number->string (cdr pair)) p)
(newline p)))

It's much more imperative, but also pretty nice and compact. The
`printf` optimization is significant for that portion of the program,
but that isn't much of the running time. The overall running time for
10 copies of the KJV is about 9 seconds on my laptop.

I think the remaining difference between Racket and other languages is
likely the `string-split` and `string-downcase` functions, plus the
relatively-inefficient string representation that Racket uses.

Sam

> To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/09c58a34-bd2d-49e7-bfbd-d3253c1e6dd1n%40googlegroups.com.

[Sam Tobin-Hochstadt]

Here are several variants of the code:
https://gist.github.com/d6fbe3757c462d5b4d1d9393b72f9ab9

The enabled version is about the fastest I can get without using
`unsafe` (which the rules said not to do). It's possible to optimize a
tiny bit more by avoiding sorting, but only a few milliseconds -- it
would be more significant if there were more different words.

Switching to bytes works correctly for the given task, but wouldn't
always work in the case of general UTF8 input. But those versions
appeared not be faster for me. Also, writing my own string-downcase
didn't help. And using a big buffer and doing my own newline splitting
didn't help either.

The version using just a regexp matching on a port (suggested by
Robby) turned out not to be faster either, so my suspicion is that the
original slowness is just using regexps for splitting words.

Sam

[Bogdan Popa]

I managed to get it about as fast as Python by making it really
imperative and rolling my own hash:

https://gist.github.com/Bogdanp/fb39d202037cdaadd55dae3d45737571

[Pawel Mosakowski]

Hi Sam,

Thank you for your responses. Your fastest approach runs in quarter of the time (~5 seconds) of my naive implementation which is pretty amazing. I already had a closer look and can see all the improvements like avoiding string-split, using a mutable hash and modifying it in place, using a vector etc. I have learned a lot and I will incorporate this in my future programs.

Many thanks,

Pawel

[Pawel Mosakowski]

Hi Bogdan,

This is a brilliant solution and also completely over my head. It finishes in ~3.75s on my PC and is faster than the Python version which basically delegates all the work to C. I will need to spend some time on understanding it but I am looking forward to learning something new.

Many thanks,

Pawel

[Sorawee Porncharoenwase]

string-split always uses regex. I wonder if a fast path when the splitter is a regular string will be worth it.

To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/ca365c42-d698-4e43-863e-7fc95849c31en%40googlegroups.com.

[Bogdan Popa]

I updated the gist with some cleanups and additional improvements that
get the runtime down to a little over 1s (vs ~350ms for the optimized C
and Rust code) on my maxed-out 2019 MBP and ~600ms on my M1 Mac Mini.

>> https://groups.google.com/d/msgid/racket-users/09c58a34-bd2d-49e7-bfbd-d3253c1e6dd1n%40googlegroups.com
>> .
>>

[Sam Tobin-Hochstadt]

One minor additional suggestion: if you use #:authentic for the struct, it will generate slightly better code for the accessors. 

Sam

To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/m2r1kb30qh.fsf%40192.168.0.142.

[Dominik Pantůček]

Another attack of [1]. But yeah, why not do some [2].

Trees to the rescue [3].

$ racket --version
Welcome to Racket v8.0 [cs].

$ racket countwords-bogdan2.rkt <kjvbible.txt |tail -1
cpu time: 135 real time: 135 gc time: 8

$ racket countwords-dzoe2.rkt <kjvbible.txt | tail -1
cpu time: 69 real time: 69 gc time: 3

I just changed (countwords) to (time (countwords)) in Bogdan's code to
measure the running time.

The difference is that I am positively defining which letters form words
(a-z, A-Z) and that all others are treated as word separators. The
buffer size is the same - and honestly, the speedup between 1024 and
1024^2 bytes buffer is barely measurable.

The only option for further speedup I can immediately think of is to
allocate a huge vector of wtnodes and change chld field to be a starting
index into this big vector (should reduce allocations).

Btw, making it unsafe does not speed it up at all (probably CS
recognizes the vectors and all those refs are inlined anyway).


Cheers,
Dominik

[1] https://xkcd.com/386/
[2] http://phdcomics.com/comics/archive.php?comicid=1735
[3] https://gist.github.com/dzoep/0e081d0544afac539a4829179c601e0e

[Bogdan Popa]

Nice! It's worth pointing out, though, that by limiting yourself to
alpha chars, you're processing about 8% less data and the results don't
pass the tests. :P

$ wc kjvbible_x10.txt
998170 8211330 43325060

$ sed 's/[a-zA-Z ]//g' < kjvbible_x10.txt | wc
998170 739310 3600800

I think your version would still be faster, but I'm curious what the
numbers would look like if only whitespace chars were considered word
separators.

[Laurent]

(Welcome to Racket v8.0.0.1 [cs]. )

All results are measured on my laptop on the 10x file with `$ time racket <file.rkt>`, thus including the Racket VM.

* Bogdan's version with #lang racket/base: 1s.

* Dominik's version with vectors of length 256 (instead of 26) and splitting on spaces/return/newline only and #lang racket/base: 1.6s

But Bogdan's printing takes only ~100ms while Dominik's takes ~400ms.

* I also implemented a variant of Dominik's 'discrimination tree' to use a hasheqv instead of a vector at each node: 6.2s (but the memory footprint is likely nicer :-p )

  This also uses `in-lines` instead of Bogdan's buffers so there may still be something to gain here.

* Replacing a hasheqv with an assoc: 4.2s.

* Starting with an assoc and switching to a hasheqv when there are too many elements (n=20, couldn't do better): 3.9s

Code is here: https://gist.github.com/Metaxal/ae0a6937d8f388f3f40ec7396041be55

I also noticed that `dict-ref` is *really* slow (35s) compared to `assv`.

@Bogdan: You can use `#:key` in `sort`.

To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/m2o8ff2vnt.fsf%40192.168.0.142.

[Sam Tobin-Hochstadt]

Use `#:authentic`, and `unsafe-vector*-{ref,set!}` saved about 50 more
ms on my machine.

Then getting rid of `set!` and just re-binding the relevant variables
produced another 50 ms speedup.

https://gist.github.com/7fc52e7bdc327fb59c8858a42258c26a

Sam

On Fri, Mar 19, 2021 at 7:21 AM Sam Tobin-Hochstadt

[Laurent]

Sam: How do you accurately measure such small speed-ups? On my machines, if I run the same program twice, I can sometimes see more than 10% time difference.

To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/CAK%3DHD%2Bbd3w-zGTC5uUUSL%3DVz97%3DkSi8WpQA%2BvcFS_0ZA9S%2BM7A%40mail.gmail.com.

[Sam Tobin-Hochstadt]

I went from numbers around 1000 ms to 950 ms to 900 ms. There was
variance around those numbers, but it was pretty consistent.

For more precise answers, there are a few things you can try. One is
to measure instructions instead of time (ie, with perf). Another is to
run it a bunch of times and take an average. The `hyperfine` tool is
good for that. But probably the best advice is to make the program
take longer so differences are more apparent -- variation usually
increases sub-linearly.

Sam

> To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/CABNTSaGuT%3DN7f6x6VGyDBrjrAohgEh7PhzecMiCwVy2ae%3DRmzw%40mail.gmail.com.

[Gustavo Massaccesi]

With two additional tricks I saved like 100ms.

* Saving the output port instead of reading the parameter implicitly each time.

* Replacing (write (cdr p)) with (write-fx cdr p)) where

(define (write-fx n [o (current-output-port)])

  ; TODO: Add negatives :)

  (if (fx> n 0)

      (let loop ([n n])
        (when (fx> n 10)
          (loop (fxquotient n 10)))
        (write-byte (fx+ 48 (fxremainder n 10)) o))

      (write-byte 48 o)))

and at the end of a program something like

(define o (current-output-port))
  (time (for ([p (in-vector items)]
        #:break (not (pair? p)))
    (write-bytes (car p) o)
    (write-byte 32 o)
    (write-fx (cdr p) o)
    (write-byte 10 o)))      ; and a closing )

Gustavo

To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/CAK%3DHD%2BYCpRmnsYCDt%3DeoVQvEz2VbLmo_JKtky9ZtxuSJB65s4Q%40mail.gmail.com.

[Bogdan Popa]

On top of these changes, replacing the hash function with Chez's
`equal-hash` saves another 50ms. That gets the runtime down to around
900ms on my machine, including Racket startup time. Ignoring startup
time, this version beats the `simple.c` implementation in the original
repo by about 50ms (800ms vs 850ms).

https://gist.github.com/Bogdanp/b9256e1a91de9083830cb616b3659ff8

>> https://groups.google.com/d/msgid/racket-users/CAK%3DHD%2Bbd3w-zGTC5uUUSL%3DVz97%3DkSi8WpQA%2BvcFS_0ZA9S%2BM7A%40mail.gmail.com

>> .
>> >
>> > --
>> > You received this message because you are subscribed to the Google
>> Groups "Racket Users" group.
>> > To unsubscribe from this group and stop receiving emails from it, send
>> an email to racket-users...@googlegroups.com.
>> > To view this discussion on the web visit

>> https://groups.google.com/d/msgid/racket-users/CABNTSaGuT%3DN7f6x6VGyDBrjrAohgEh7PhzecMiCwVy2ae%3DRmzw%40mail.gmail.com

>> .
>>
>> --
>> You received this message because you are subscribed to the Google Groups
>> "Racket Users" group.
>> To unsubscribe from this group and stop receiving emails from it, send an
>> email to racket-users...@googlegroups.com.
>> To view this discussion on the web visit

>> https://groups.google.com/d/msgid/racket-users/CAK%3DHD%2BYCpRmnsYCDt%3DeoVQvEz2VbLmo_JKtky9ZtxuSJB65s4Q%40mail.gmail.com
>> .
>>