The overhead of abstraction in C/C++ vs. Python/Ruby

http://blog.reverberate.org/2014/10/the-overhead-of-abstraction-in-cc-vs.html

by frostmatthew • 12 years ago

51 40 12 years ago

blog.reverberate.org

TillE• 12 years ago

The "C with classes" code looks so weird. I'm a little surprised you got away with int main() not explicitly returning 0, didn't realize that was still valid C++. Anyway, C doesn't really do abstractions to any great extent, so it would make more sense to focus only on C++ and delve a little deeper.

Observe what happens when you make a method virtual, for example.

jeorgun• 12 years ago

V-tables aren't a very expensive abstraction. They prevent inlining, which isn't exactly ideal (hell, that's the main point of the post), but nowadays their cost over a normal function call is basically negligible.

coherentpony• 12 years ago

V-tables aren't a very expensive abstraction. They prevent inlining, which isn't exactly ideal (hell, that's the main point of the post), but nowadays their cost over a normal function call is basically negligible.

This is not true. Sweeping general statements like this are almost never true. It is really problem dependent whether a particular type of abstraction is 'too expensive' or not.

I have to actively avoid vtbl lookups in my domain, and this is because if I don't I lose about 15% in execute time. You might think 15% "isn't very long", but it is when your application uses several thousand MPI processes.

StephanTLavavej• 12 years ago

C++ did that first. C++98 and C99 implicitly "return 0;" when falling off the end of main(). It was C89 that required "return 0;" here.

TheLoneWolfling2• 12 years ago

Read: "The overhead of abstraction in languages that do optimizing compilation versus languages that don't".

haberman7• 12 years ago

Also: CPython and MRI don't do optimizing compilation, which is not entirely obvious considering they both do compile to bytecode and optimize slightly.

TheLoneWolfling2• 12 years ago

Don't know about MRI. With Python (as it's the reference implementation, I don't see the need to call it by another name) all, IIRC, compilation does with optimization enabled is remove assert and if __debug__: statements. Oh, and docstrings. So yeah... slightly. To put it mildly.

(Note that with this "benchmark", PyPy has (almost) no difference in runtime between the abstracted and non-abstracted versions.)

mackwic• 12 years ago

Even though the question is good, the article answer is disappointing.

First, I really don't like the angle taken, then the question of abstraction (why we do it, how) and choices made by languages designers (and variations in their idioms) are so vast that you really can't treat the question through a <1000 signs blog post.

Some quick points:

- You don't code for the machine, you code to be read by another human (possibly you) in the future. I insist, you will be read regularly and frequently. Thus, your code needs to be clear, precise, and concise. This must be the first thing in mind when coding: program what need to be done in a way that a fellow stranger could understand.

- Abstraction is a way to keep a structure of code clear when the interactions become complex and/or abundant. If you can avoid them when still being crystal clear in your code, do it. Direct speaking is always better than convolution.

- The main requirements when you code are often one or two of: quick to develop, easy to maintain, extensible, efficient (you control your big-O and _WORST_ exec times), correct (no bug. at. all.), real time (when X happens, Y is done between n µs and m µs during p ns)

- So, know when performance is a goal, and know when it's not. Choose your language, your techs, your team considering these goals.

- And yeah, C89, C99 and the C++es have a very high overhead of abstraction: clarity, concision, and sometimes performance (all abstractions can't be inlined). Think of it.

haberman7• 12 years ago

You are speaking in vague platitudes, most of which don't contradict what I wrote in the article at all.

I was speaking of a specific context in which I was working (writing C or C++ extensions for Python and Ruby), and addressing a specific design question (how much of the library should be in C or C++ vs. Python or Ruby). I guess I didn't specifically say this, but I thought I made it pretty obvious that I'm working in a situation where performance is a factor.

My observations were also specific to the question of the relative cost of creating layering abstractions in Python/Ruby vs. C/C++.

And yeah, C89, C99 and the C++es have a very high overhead of abstraction: clarity, concision

Now you're just being silly.

Abstractions are sometimes good, making things clearer and more concise. The wrong abstractions can make code less clear and concise.

Here is my solution to that problem: if an abstraction is not making your code clearer and more concise, don't use it. Saying that abstractions in C and C++ are inherently less clear and concise just shows anti-C/C++ bias.

theflubba3• 12 years ago

I don't know if your solution is correct. Many functional abstractions in Haskell, such as those that involve high level typeclass abstractions, make code more esoteric, and not more clear. However this is a tradeoff to make code more based in algebra and logical structure. This is more important than readability and conciseness.

adamnemecek• 12 years ago

You don't code for the machine, you code to be read by another human

This is like the new 'premature optimization is the root of all evil'. Mindlessly repeated without considering any context.

userbinator• 12 years ago

You don't code for the machine, you code to be read by another human (possibly you) in the future.

This is one of those sayings (along with the one on "premature optimisation") that I believe has caused huge amounts of resource wastage over the years, and made much software orders of magnitude more inefficient that it could be. Only the exponential growth in hardware performance has hidden that waste, but that is (fortunately?) coming to an end.

Despite humans writing it, code does not exist mainly to be read by humans. It exists to be executed by a machine, to solve a real problem for its users. Your users do not care about your source code. They want a small, fast and efficient application that does what they want. The majority of useful code is executed far more than it's read or written.

program what need to be done in a way that a fellow stranger could understand

This is vague. I agree with you that writing purposely obtuse code for no other reason isn't a good idea, but you should also be exploiting all the power your programming language has to offer when it can make your code simpler and more efficient. I believe that if someone else using the same programming language cannot understand your code, then they should improve themselves so they can.

RedNifre• 12 years ago

No, you code to get the job done in an optimal way which usually means for the minimum amount of cost.

Given that the cost for a developer is roughly constant while computation time gets cheaper, this means that nowadays it's best to write code that is quick to modify and thus cheap to modify. The way to do this is to make the code easy to read.

(There are exceptions where you end up writing difficult Fortran code on a supercomputer, but usually "You don't code for the machine, you code to be read by another human (possibly you) in the future." is a good rule of thumb for efficiency.)

jerven• 12 years ago

Maybe, this person should understand what they are timing. The time command is a wonderful tool, but comparing the execution time of a 4 line C program with the start up cost of a VM and interpreter is inane.

One needs to measure the program doing the real work. And in this case for the C program you might find that bash forking is the actual cost not the C code running through the loop.

JRuby and PyPy are starting to do things with their specialisation tricks that mean that the overhead of abstraction is 1000 loops before it goes away.

A performance benchmark running for less than a few minutes is a joke. Startup variance alone will dominate the results.

Chris911• 12 years ago

Except he's not comparing the C program and the Python one. He's comparing the time difference (or ratio) for each language with and without abstraction.

haberman7• 12 years ago

comparing the execution time of a 4 line C program with the start up cost of a VM and interpreter is inane.

That's not what I compared. Maybe you should read the article again?

And in this case for the C program you might find that bash forking is the actual cost not the C code running through the loop.

I didn't publish any benchmark numbers of any C programs, because I didn't need to: the two C++ programs I was comparing compiled to exactly the same machine code, making empirical observations of their execution time immaterial.

theflubba3• 12 years ago

This article conflates abstraction with OO and language features, IMO.

Putting some functionality into an object is not as interesting as running that functionality as a monadic command.

Would be more interesting to run some benchmarks in Haskell using functional abstractions rather than OO machinery.

jaimebuelta• 12 years ago

The problem with this article is that is doing something extremely simple, and calling that an abstraction.

The power of Python or Ruby is not demonstrated doing a for loop, is with something like this:

print sum(int(n) for n in open('/a/file/with/numbers'))

We don't need to care about a lot of stuff that we would in C/C++ (memory management, iterate through a file, the format of stuff, terminating the loop properly)

Are C/C++ interesting languages? Of course, sometimes you need that extra level of control about what's going on.

But I don't agree that the cost of abstraction is higher in Python/Ruby. You can create code that does A LOT in a few lines, that most of the times will have a good enough speed.

In C/C++ you need a lot of work to create a good abstraction, that's not guaranteed to perform well.

vinkelhake• 12 years ago

It's not quite that bad. To sum the numbers in a file you've got a line to open the file and another to call std::accumulate with istream iterators. Memory management? Doesn't come up. Terminating the loop? What loop?

ajtulloch• 12 years ago

FWIW, with folly::gen [1], this is just:

    using namespace folly::gen;
    byLine('/a/file/with/numbers') | eachTo<int> | sum;

https://github.com/facebook/folly/tree/master/folly/gen

seri• 12 years ago

Thank you for this. Got me really interested in folly now.

nly• 12 years ago

It's only 2 lines even restricting yourself to the standard library

    #include <fstream>
    #include <iterator>
    #include <numeric>
    using namespace std;

    int main() {
        ifstream nf ("numbers.txt");
        return accumulate (istream_iterator<int>(nf), istream_iterator<int>(), 0);
    }

jeffdavis• 12 years ago

I think the post would be more clear if the author said "indirection" or "path length" (loosely: the number of primitive logical steps to complete an operation).

The point is that, with many interpreted language compilers, you doubly lose out: the primitive logical steps (in the author's example, function calls) are slower, and they aren't optimized away like they are in C.

I think that's a valid point, because it limits code reuse. As you build up vast libraries and frameworks, the trivial indirections build up and it slows everything to a crawl.

frostmatthew• 12 years ago

But I don't agree that the cost of abstraction is higher in Python/Ruby. You can create code that does A LOT in a few lines, that most of the times will have a good enough speed.

I think you missed the author's point - you may want to re-read the blog's first two paragraphs.

haberman7• 12 years ago

The problem with this article is that is doing something extremely simple, and calling that an abstraction.

Simple examples make for good blog articles. Real abstractions follow the same pattern as simple ones.

The power of Python or Ruby is not demonstrated doing a for loop, is with something like this: print sum(int(n) for n in open('/a/file/with/numbers'))

In this case, your abstraction is sum(). Internally, sum() will use something like a for loop in its implementation. sum() is fast because it's implemented in C, not Python.

zem• 12 years ago

i was expecting just the opposite - the overhead in terms of programmer time and effort involved in abstracting something out versus just copy/pasting the code.

tokenrove• 12 years ago

It would be interesting to see the same comparison with a whole-program optimizing compiler like MLton, especially with the abstraction split across source file boundaries (an area where C compilers have gotten much stronger lately).

chrisseaton• 12 years ago

A while ago I wrote a blog post about how compilers for Ruby can remove abstraction. Features that you might think of as being expensive abstractions such as using #send to call methods instead of calling them directly, need not have any overhead at all with a good enough compiler http://www.chrisseaton.com/rubytruffle/pushing-pixels/

sbov• 12 years ago

The overhead of abstraction is slightly more than 1.5 addition operations? Sign me up.

astrobe_• 12 years ago

I think that this observation is not as relevant to the task at hand that it seems. The author is in the same position as the implementor of a VM, who has to decide which operation will be a primitive and which will not be. What is actually more important than anything else if you care about performance is which operations are likely to be called from hot loops. It's classic optimization wisdom except that you can only work on expected scenarios rather than real-world uses (until you release your stuff and other people use it). So a 100x slower abstraction is ok(-ish) as long as it's not somthing to be used in performance critical sections.

haberman7• 12 years ago

Your comment amounts to: "this performance-related observation doesn't apply in situations where performance is not an issue."

Yes, but I am working in a situation where performance is an issue. So yes, the observation is relevant to the task at hand.

The overhead of abstraction in C/C++ vs. Python/Ruby

Related Stories