3 ms·

> If you're using C or C++ "for performance" and not using a profiler in day-to-day development, you don't need to be using C or C++ and should be using a memor

• lmm 6y ago
  > If you are making building blocks (libraries) for other languages, you have to use C or C++ (maybe Rust soon). Sure (though there are often other ways to achieve the thing you actually want to do). But in that case you're not doing it "for performance". > When you aim for high performance you don't profile day 1, that's a complete waste of time. You will never transform every of your function in a critical kernel. You profile when you reach a performance bottleneck and you need it. In that case your performance requirements are not extreme enough to justify using C/C++ for your whole application. Write it in a safer language, when you hit performance issues profile and optimise, and maybe drop into C/C++ for those few "critical kernels" in the unlikely event that it turns out you actually need to.
  • adev_ 6y ago
    > In that case your performance requirements are not extreme enough to justify using C/C++ for your whole application. Again : no. That is in my experience both wrong and over-idealist. For many applications, the overhead in term of dev time of writing bindings for every of your compute kernels + the pleasure of debugging the problems associated with them and heterogeneous build chains is generally several order of magnitude higher in term of man-hour-cost than just doing your program entirely in C/C++/rust. There is an other aspects generally ignored: - Theory say that often 80% of the compute time is consumed in 20% of the code. That's often wrong, and many HPC simulators do not have any kernel taking more than 3-7% of total run time. Consequently, everything might one day need to be optimised. - Many performance critical algorithm are state of art and evolve. Meaning your innocent little function in "memory managed" language might become tomorrow a new bottleneck. And you do not want to have to rewrite that all the time. A lot of new devs are wrongly scared of manual memory management where it became a non-problem with RAII in C++[1-2]x or the borrow checker in Rust. And generally the ones that are scared are the one that do not use it. The mental overhead with memory in C++ does not come so much with the lifetime of object, it comes mainly with HOW to use efficiently YOUR memory: aligned object in memory, cache effect, indirection, cost of polymorphism, allocation, etc, etc. All these aspects, you do not think about them in memory managed language because you can not: You have no control over it. And that's also why they bite you in the face in term of performance, generally much more than the 2x you quoted before. Just a remember, a cache miss and it's ~200 cycle you loose
    • lmm 6y ago
      > - Theory say that often 80% of the compute time is consumed in 20% of the code. That's often wrong, and many HPC simulators do not have any kernel taking more than 3-7% of total run time. Consequently, everything might one day need to be optimised. > - Many performance critical algorithm are state of art and evolve. Meaning your innocent little function in "memory managed" language might become tomorrow a new bottleneck. And you do not want to have to rewrite that all the time. You can't have it both ways. If it's really common for everything to become a performance bottleneck, it's worth profiling from the start so that you avoid having major pitfalls anywhere. If it's rare and exceptional, FFI for those cases is fine. > The mental overhead with memory in C++ does not come so much with the lifetime of object, it comes mainly with HOW to use efficiently YOUR memory: aligned object in memory, cache effect, indirection, cost of polymorphism, allocation, etc, etc. > All these aspects, you do not think about them in memory managed language because you can not: You have no control over it. > And that's also why they bite you in the face in term of performance, generally much more than the 2x you quoted before. Just a remember, a cache miss and it's ~200 cycle you loose On the contrary. Plenty of people do those kind of things in, say, Java. They require knowing about compiler internals and using unsupported hints, or even bypassing parts of the compiler. But so does controlling these things in in C++.
      • adev_ 6y ago
        > If it's really common for everything to become a performance bottleneck, it's worth profiling from the start so that you avoid having major pitfalls anywhere. If it's rare and exceptional, FFI for those cases is fine. You can have it both way, code evolve. It is pretty common in performance critical code that a minor, almost never call function in one scenario, become a performance critical bottleneck in an other. If you already played with large scaled simulation software, this happened almost every week depending of your inputs on what you are interested to simulate. > On the contrary. Plenty of people do those kind of things in, say, Java. They require knowing about compiler internals and using unsupported hints, or even bypassing parts of the compiler No it's not again. I have been developing in C++ for 15 years, including in the HPC world and I (close to) never had to touch a compiler internal. The language gives you what you need for performances, you do not need to play with that. At the opposite JIT compilers like V8 or Java are monster of complexity very sensible to side effect [1] and controlling things like "Does this data fit in my L2 cache ?" in them is close to impossible because even things like "Where are my data and what are there size ?" is an hard question. Once again, there is theory and there is practice. Theory is what you say. Practice is that 98% of performance critical software in HPC, game industry, physics and High Frequency Trading is in C++/C (maybe Rust soon). And this is why. [1]: https://mathiasbynens.be/notes/prototypes https://mathiasbynens.be/notes/prototypes