Many of these have C-bindings for their libraries, which means that slowness is caused by bad code (such as making a for loop with a C-call for each iteration instead of once for the whole loop).
I am no coder, but it is my experience that bad code can be slow regardless of language used.
Bad code can certainly be part of it. The average skill level of those coding C/C++/Rust tends to be higher. And modern programs typically use hundreds of libraries, so even if your own code is immaculate, not all of your dependencies will be.
But there’s other reasons, too:
Python, Java etc. execute their compiler/interpreter while the program is running.
CLIs are magnitudes slower, because these languages require a runtime to be launched before executing the CLI logic.
GUIs and simulations stutter around, because these languages use garbage collection for memory management.
And then just death by a thousand paper cuts. For example, when iterating over text, you can’t tell it to just give you a view/pointer into the existing memory of the text. Instead, it copies each snippet of text you want to process into new memory.
And when working with multiple threads in Java, it is considered best practice to always clone memory of basically anything you touch. Like, that’s good code and its performance will be mediocre. Also, you better don’t think about using multiple threads in Python+JS. For those two, even parallelism was an afterthought.
Well, and then all of the above feeds back into all the libraries not being performant. There’s no chance to use the languages for performance-critical stuff, so no one bothers optimizing the libraries.
Idk numpy go brrrrrrrrrr. I think it’s more just the right tool for the right job. Most languages have areas they excel at, and areas where they’re weaker, siloing yourself into one and thinking it’s faster for every implementation seems short sighted.
At it’s heart, numpy is C tho. That’s exactly what I’m talking about. Python is amazing glue code. It makes this fast code more useful by wrapping it in simple® scripts and classes.
For example, when iterating over text, you can’t tell it to just give you a view/pointer into the existing memory of the text. Instead, it copies each snippet of text you want to process into new memory.
As someone used to embedded programming, this sounds horrific.
Yep. I used to code a lot in JVM languages, then started learning Rust. My initial reaction was “Why the hell does Rust have two string types?”.
Then I learned that it’s for representing actual memory vs. view and what that meant. Since then I’m thinking “Why the hell do JVM languages not have two string types?”.
I’m not a java programmer, but I think the equivalent to str would be char[]. However the ergonomics of rust for str isn’t there for char[], so java devs probably use String everywhere.
Many of these have C-bindings for their libraries, which means that slowness is caused by bad code (such as making a for loop with a C-call for each iteration instead of once for the whole loop).
I am no coder, but it is my experience that bad code can be slow regardless of language used.
Bad code can certainly be part of it. The average skill level of those coding C/C++/Rust tends to be higher. And modern programs typically use hundreds of libraries, so even if your own code is immaculate, not all of your dependencies will be.
But there’s other reasons, too:
And when working with multiple threads in Java, it is considered best practice to always clone memory of basically anything you touch. Like, that’s good code and its performance will be mediocre. Also, you better don’t think about using multiple threads in Python+JS. For those two, even parallelism was an afterthought.
Well, and then all of the above feeds back into all the libraries not being performant. There’s no chance to use the languages for performance-critical stuff, so no one bothers optimizing the libraries.
Java is still significantly faster and more efficient than Python tho - because it has ahead-of-time optimizations and is not executing plain text.
Idk numpy go brrrrrrrrrr. I think it’s more just the right tool for the right job. Most languages have areas they excel at, and areas where they’re weaker, siloing yourself into one and thinking it’s faster for every implementation seems short sighted.
At it’s heart, numpy is C tho. That’s exactly what I’m talking about. Python is amazing glue code. It makes this fast code more useful by wrapping it in simple® scripts and classes.
Python is the slowest (widely used) language there is. It’s not hard to be faster.
As someone used to embedded programming, this sounds horrific.
Yep. I used to code a lot in JVM languages, then started learning Rust. My initial reaction was “Why the hell does Rust have two string types?”.
Then I learned that it’s for representing actual memory vs. view and what that meant. Since then I’m thinking “Why the hell do JVM languages not have two string types?”.
I’m not a java programmer, but I think the equivalent to str would be char[]. However the ergonomics of rust for str isn’t there for char[], so java devs probably use String everywhere.
At least with Java, its the over(ab)use of Reflections and stuff like dependency injection that slows things down to a crawl.