By David Copernicus Bowie


2019-04-14 12:07:29 8 Comments

I am following the Rust track on Exercism.io. I have a fair amount of C/C++ experience. I like the 'functional' elements of Rust but I'm concerned about the relative performance.

I solved the 'run length encoding' problem:

pub fn encode(source: &str) -> String {
    let mut retval = String::new();
    let firstchar = source.chars().next();
    let mut currentchar = match firstchar {
        Some(x) => x,
        None => return retval,
    };
    let mut currentcharcount: u32 = 0;
    for c in source.chars() {
        if c == currentchar {
            currentcharcount += 1;
        } else {
            if currentcharcount > 1 {
                retval.push_str(&currentcharcount.to_string());
            }
            retval.push(currentchar);
            currentchar = c;
            currentcharcount = 1;
        }
    }
    if currentcharcount > 1 {
        retval.push_str(&currentcharcount.to_string());
    }
    retval.push(currentchar);
    retval
}

I noticed that one of the top-rated answers looked more like this:

extern crate itertools;

use itertools::Itertools;

pub fn encode(data: &str) -> String {
    data.chars()
        .group_by(|&c| c)
        .into_iter()
        .map(|(c, group)| match group.count() {
            1 => c.to_string(),
            n => format!("{}{}", n, c),
        })
        .collect()
}

I love the top rated solution; it is simple, functional, and elegant. This is what they promised me Rust would be all about. Mine on the other hand is gross and full of mutable variables. You can tell I'm used to C++.

My problem is that the functional style has a SIGNIFICANT performance impact. I tested both versions with the same 4MB of random data encoded 1000 times. My imperative solution took under 10 seconds; the functional solution was ~2mins30seconds.

  • Why is the functional style so much slower than the imperative style?
  • Is there some problem with the functional implementation which is causing such a huge slowdown?
  • If I want to write high performance code, should I ever use this functional style?

2 comments

@Shepmaster 2019-04-14 14:49:52

TL;DR

A functional implementation can be faster than your original procedural implementation, in certain cases.

Why is the functional style so much slower than the imperative style? Is there some problem with the functional implementation which is causing such a huge slowdown?

As Matthieu M. already pointed out, the important thing to note is that the algorithm matters. How that algorithm is expressed (procedural, imperative, object-oriented, functional, declarative) generally doesn't matter.

I see two main issues with the functional code:

  • Allocating numerous strings over and over is inefficient. In the original functional implementation, this is done via to_string and format!.

  • There's the overhead of using group_by, which exists to give a nested iterator, which you don't need just to get the counts.

Using more of itertools (batching, take_while_ref, format_with) brings the two implementations much closer:

pub fn encode_slim(data: &str) -> String {
    data.chars()
        .batching(|it| {
            it.next()
                .map(|v| (v, it.take_while_ref(|&v2| v2 == v).count() + 1))
        })
        .format_with("", |(c, count), f| match count {
            1 => f(&c),
            n => f(&format_args!("{}{}", n, c)),
        })
        .to_string()
}

A benchmark of 4MiB of random alphanumeric data, compiled with RUSTFLAGS='-C target-cpu=native':

encode (procedural)     time:   [21.082 ms 21.620 ms 22.211 ms]

encode (fast)           time:   [26.457 ms 27.104 ms 27.882 ms]
Found 7 outliers among 100 measurements (7.00%)
  4 (4.00%) high mild
  3 (3.00%) high severe

If you are interested in creating your own iterator, you can mix-and-match the procedural code with more functional code:

struct RunLength<I> {
    iter: I,
    saved: Option<char>,
}

impl<I> RunLength<I>
where
    I: Iterator<Item = char>,
{
    fn new(mut iter: I) -> Self {
        let saved = iter.next(); // See footnote 1
        Self { iter, saved }
    }
}

impl<I> Iterator for RunLength<I>
where
    I: Iterator<Item = char>,
{
    type Item = (char, usize);

    fn next(&mut self) -> Option<Self::Item> {
        let c = self.saved.take().or_else(|| self.iter.next())?;

        let mut count = 1;
        while let Some(n) = self.iter.next() {
            if n == c {
                count += 1
            } else {
                self.saved = Some(n);
                break;
            }
        }

        Some((c, count))
    }
}

pub fn encode_tiny(data: &str) -> String {
    use std::fmt::Write;

    RunLength::new(data.chars()).fold(String::new(), |mut s, (c, count)| {
        match count {
            1 => s.push(c),
            n => write!(&mut s, "{}{}", n, c).unwrap(),
        }
        s
    })
}

1 — thanks to Stargateur for pointing out that eagerly getting the first value helps branch prediction.

A benchmark of 4MiB of random alphanumeric data, compiled with RUSTFLAGS='-C target-cpu=native':

encode (procedural)     time:   [19.888 ms 20.301 ms 20.794 ms]
Found 4 outliers among 100 measurements (4.00%)
  3 (3.00%) high mild
  1 (1.00%) high severe

encode (tiny)           time:   [19.150 ms 19.262 ms 19.399 ms]
Found 11 outliers among 100 measurements (11.00%)
  5 (5.00%) high mild
  6 (6.00%) high severe

I believe this more clearly shows the main fundamental difference between the two implementations: an iterator-based solution is resumable. Every time we call next, we need to see if there was a previous character that we've read (self.saved). This adds a branch to the code that isn't there in the procedural code.

On the flip side, the iterator-based solution is more flexible — we can now compose all sorts of transformations on the data, or write directly to a file instead of a String, etc. The custom iterator can be extended to operate on a generic type instead of char as well, making it very flexible.

See also:

If I want to write high performance code, should I ever use this functional style?

I would, until benchmarking shows that it's the bottleneck. Then evaluate why it's the bottleneck.

Supporting code

Always got to show your work, right?

benchmark.rs

use criterion::{criterion_group, criterion_main, Criterion}; // 0.2.11
use rle::*;

fn criterion_benchmark(c: &mut Criterion) {
    let data = rand_data(4 * 1024 * 1024);

    c.bench_function("encode (procedural)", {
        let data = data.clone();
        move |b| b.iter(|| encode_proc(&data))
    });

    c.bench_function("encode (functional)", {
        let data = data.clone();
        move |b| b.iter(|| encode_iter(&data))
    });

    c.bench_function("encode (fast)", {
        let data = data.clone();
        move |b| b.iter(|| encode_slim(&data))
    });

    c.bench_function("encode (tiny)", {
        let data = data.clone();
        move |b| b.iter(|| encode_tiny(&data))
    });
}

criterion_group!(benches, criterion_benchmark);
criterion_main!(benches);

lib.rs

use itertools::Itertools; // 0.8.0
use rand; // 0.6.5

pub fn rand_data(len: usize) -> String {
    use rand::distributions::{Alphanumeric, Distribution};
    let mut rng = rand::thread_rng();
    Alphanumeric.sample_iter(&mut rng).take(len).collect()
}

pub fn encode_proc(source: &str) -> String {
    let mut retval = String::new();
    let firstchar = source.chars().next();
    let mut currentchar = match firstchar {
        Some(x) => x,
        None => return retval,
    };
    let mut currentcharcount: u32 = 0;
    for c in source.chars() {
        if c == currentchar {
            currentcharcount += 1;
        } else {
            if currentcharcount > 1 {
                retval.push_str(&currentcharcount.to_string());
            }
            retval.push(currentchar);
            currentchar = c;
            currentcharcount = 1;
        }
    }
    if currentcharcount > 1 {
        retval.push_str(&currentcharcount.to_string());
    }
    retval.push(currentchar);
    retval
}

pub fn encode_iter(data: &str) -> String {
    data.chars()
        .group_by(|&c| c)
        .into_iter()
        .map(|(c, group)| match group.count() {
            1 => c.to_string(),
            n => format!("{}{}", n, c),
        })
        .collect()
}

pub fn encode_slim(data: &str) -> String {
    data.chars()
        .batching(|it| {
            it.next()
                .map(|v| (v, it.take_while_ref(|&v2| v2 == v).count() + 1))
        })
        .format_with("", |(c, count), f| match count {
            1 => f(&c),
            n => f(&format_args!("{}{}", n, c)),
        })
        .to_string()
}

struct RunLength<I> {
    iter: I,
    saved: Option<char>,
}

impl<I> RunLength<I>
where
    I: Iterator<Item = char>,
{
    fn new(mut iter: I) -> Self {
        let saved = iter.next();
        Self { iter, saved }
    }
}

impl<I> Iterator for RunLength<I>
where
    I: Iterator<Item = char>,
{
    type Item = (char, usize);

    fn next(&mut self) -> Option<Self::Item> {
        let c = self.saved.take().or_else(|| self.iter.next())?;

        let mut count = 1;
        while let Some(n) = self.iter.next() {
            if n == c {
                count += 1
            } else {
                self.saved = Some(n);
                break;
            }
        }

        Some((c, count))
    }
}

pub fn encode_tiny(data: &str) -> String {
    use std::fmt::Write;

    RunLength::new(data.chars()).fold(String::new(), |mut s, (c, count)| {
        match count {
            1 => s.push(c),
            n => write!(&mut s, "{}{}", n, c).unwrap(),
        }
        s
    })
}

#[cfg(test)]
mod test {
    use super::*;

    #[test]
    fn all_the_same() {
        let data = rand_data(1024);

        let a = encode_proc(&data);
        let b = encode_iter(&data);
        let c = encode_slim(&data);
        let d = encode_tiny(&data);

        assert_eq!(a, b);
        assert_eq!(a, c);
        assert_eq!(a, d);
    }
}

@Matthieu M. 2019-04-14 17:39:05

Great iterator!

@Matthieu M. 2019-04-15 08:16:35

By the way, when resuming an iterator introduces a branch, it's possible to implement try_fold directly rather than relying on the default implementation (which calls next). This helps when the optimizer fails to optimize out the branch.

@Matthieu M. 2019-04-14 12:42:52

Let's review the functional implementation!

Memory Allocations

One of the big issues of the functional style proposed here is the closure passed to the map method which allocates a lot. Every single character is first mapped to a String before being collected.

It also uses the format machinery, which is known to be relatively slow.

Sometimes, people try way too hard to get a "pure" functional solution, instead:

let mut result = String::new();
for (c, group) in &source.chars().group_by(|&c| c) {
    let count = group.count();
    if count > 1 {
        result.push_str(&count.to_string());
    }

    result.push(c);
}

is about as verbose, yet only allocates when count > 1 just like your solution does and does not use the format machinery either.

I would expect a significant performance win compared to the full functional solution, while at the same time still leveraging group_by for extra readability compared to the full imperative solution. Sometimes, you ought to mix and match!

@David Copernicus Bowie 2019-04-14 13:55:14

That certainly gives us a speed boost, but it is still around 3x slower than the imperative version (30s rather than 10s in my tests). In fact, even if I only push a constant letter in that for loop it is still about 14s, so around 50% slower than the imperative version. That leads me to believe that group_by is probably not zero cost for this use case. Answer accepted anyway!

Related Questions

Sponsored Content

15 Answered Questions

[SOLVED] What is 'Currying'?

27 Answered Questions

[SOLVED] What is tail recursion?

30 Answered Questions

[SOLVED] map function for objects (instead of arrays)

45 Answered Questions

[SOLVED] What is a monad?

11 Answered Questions

24 Answered Questions

6 Answered Questions

16 Answered Questions

18 Answered Questions

4 Answered Questions

[SOLVED] Functional programming vs Object Oriented programming

Sponsored Content