During all that time I am engaged with programming, I hear that C and C++ are the speed standards. Fastest of the fastest, compiled straight to assembly code, nothing may compete in speed with C or C++. And, nobody seem to challenge that common belief.
Computing performance
Arithmetic operations with numbers, obviously, must work significantly faster in C than in any other language. But do they?
Some time ago I decided to write a set of simple benchmarks for many different languages to see, how large difference in speed really is.
Idea was simple : to find the sum of the one billion integer numbers, starting from zero, using straight-forward computing. Some compilers (rustc, for example) replace such simple cycles with formula expression, which, of course, will be evaluated in the constant time. To avoid that with such compilers. I used similar in costs operations with numbers, such as bitwise or.
After I got results, I was surprised very much. My world view turned upside down, and I had to reconsider everything I knew about speed of programming languages.
You may see my results in the table below :
Linux 64bit, 1.1 GHz CPU, 4GB RAM
| Language | compiler/version/args | time |
|---|---|---|
| Rust (bitwise or instead of +) | rustc 1.75.0 with -O3 | 167 ms |
| C | gcc 11.4.0 with -O3 | 335 ms |
| NASM | 2.15.05 | 339 ms |
| Go | 1.18.1 | 340 ms |
| Java | 17.0.13 | 345 ms |
| Clojure | 1.10.2 | 373 ms |
| Common Lisp | SBCL 2.1.11 | 1 sec |
| Python 3 | pypy 3.8.13 | 1.6 sec |
| Python 3 | cpython 3.10.12 | 26 sec |
| Ruby | 3.0.2p107 | 38 sec |
All tests sources you may find here :
https://github.com/Taqmuraz/speed-table
So, as we may see, C is not very much faster than Java, difference is about 3%. Also, we see that other compiled languages are very close in arithmetic operations performance to C (Rust is even faster). Dynamic languages, compiled with JIT compiler, show worse results -- mostly because arithmetic operations are wrapped into dynamically dispatched functions there.
Interpreted dynamic languages with no JIT compiler show worst performance, not a surprise.
Memory allocation performance
After that crushing defeat, C fans would say that memory allocation in C is very much faster, because you are allocating it straight from the system, not asking GC.
Now and after I will use GC term both as garbage collector and as managed heap, depending on the context.
So, why people think, that GC is so slow? In fact, GC has pre-allocated memory, and allocation is simply moving pointer to the right. Mostly GC fills allocated memory by zeros using system call, similar to memset from C, so it takes constant time. While memory allocation in C takes undefined time, because it depends on the system and already allocated memory.
But, even considering this knowledge, I could not expect so good results from Java, which you may see in following tables :
| 1.1 GHz 2 cores, 4 GB RAM |
| Running tests on single thread. |
| Result format : "Xms-Yms ~Z ms" means tests took from X to Y milliseconds, and Z milliseconds in average |
Allocating integer arrays
| integers array size | times | Java 17.0.13 new[] | C gcc 11.4.0 malloc | Common Lisp SBCL 2.1.11 make-array |
|---|---|---|---|---|
| 16 | 10000 | 0-1ms, ~0.9ms | 1-2ms, ~1.2ms | 0-4ms, ~0.73ms |
| 32 | 10000 | 1-3ms, ~1.7ms | 1-3ms, ~1.7ms | 0-8ms, ~2.ms |
| 1024 | 10000 | 6-26ms, ~12ms | 21-46ms, ~26ms | 12-40ms, ~7ms |
| 2048 | 10000 | 9-53ms, ~22ms | 24-52ms, ~28ms | 12-40ms, ~19ms |
| 16 | 100000 | 0-9ms, ~2ms | 6-23ms, ~9ms | 4-24ms, ~7ms |
| 32 | 100000 | 0-14ms, ~3ms | 10-15ms, ~11ms | 3-8ms, ~7ms |
| 1024 | 100000 | 0-113ms, ~16ms | 234-1156ms, ~654ms | 147-183ms, ~155ms |
| 2048 | 100000 | 0-223ms, ~26ms | 216-1376ms, ~568ms | 299-339ms, ~307ms |
Allocating instance of the class Person with one integer field.
| how many instances | Java 17.0.3 new Person(n) | C++ g++ 11.4.0 new Person(n) |
|---|---|---|
| 100000 | 0-6ms, ~1.3ms | 4-8ms, ~5ms |
| 1 million | 0-11ms, ~2ms | 43-69ms, ~47ms |
| 1 billion | 22-50ms, ~28ms | process terminated |
All tests sources you may find here :
https://github.com/Taqmuraz/alloc-table
There I tested four languages in total : C, C++, Java and Lisp. And, languages with GC always show better results, though I tested them much stricter, than C and C++. For example, in Java I am allocating memory through the virtual function call, so it may not be statically optimized, and in Lisp I am checking first element of the allocated array, so compiler won't skip allocation call.
Releasing memory
C fans are still motivated to protect their beliefs, so, they say "Yes, you do allocate memory faster, but you have to release it after!".
True. And, suddenly, GC releases memory faster, than C. But how? Imagine, we made 1 million allocations from GC, but later we have only 1000 objects referenced in our program. And, let's say, those objects are distributed through all of that long span of memory. GC does stack tracing, finding those 1000 "alive" objects, moves them to the previous generation heap peak and puts heap peak pointer after the last of them. That's all.
So, no matter, how many objects you allocate, GC's work time is decided by how many of them you keep after.
And, in opposite with that, in C you have to release all allocated memory manually, so, if you allocated memory 1 million times, you have to make 1 million release calls as well (or you are going to have memory leaks). That means, O(1)-O(n) of GC against O(n) or worse of C, where n is the number of allocations happened before.
Summary
So, I want to consolidate the victory of garbage collected languages over C and C++. Here is the summary table :
| demands | languages with GC | C/C++ |
|---|---|---|
| arithmetic | fast with JIT | fast |
| allocating memory | fast O(1) | slow |
| releasing memory | fast O(1) best case, O(n) worst case | O(n) or slower |
| memory safe | yes | no |
Now we may see -- garbage collection is not a necessary evil, but the best thing we could only wish to have. It gives us safety and performance both.
Tribute to C
While C does show worse results on my tests, it is still an important language and it has own application field. My article does not aim for C rejection or obliteration. C is not bad, it is just not that superior as people think. Many good projects collapsed only because some people decided to use C instead of Java, for example, because they have been told that C is very much faster, and Java is incredibly slow because of garbage collection. C is good, when we write very small and simple programs. But, I would never recommend writing complex programs or games with C.
C++ is different
C++ is not simple, is not flexible, has overloaded syntax and too much complicated specification. Programming with C++ you will not implement own ideas but fight with compiler and memory errors 90% of the time.
This article does aim for rejection of C++, because speed and performance are only excuses people give for using this language in software development. Using C++, you are paying with your time, your program performance and your mental health. So, when you have choice between C++ and any other language, I hope you choose the last one.
Top comments (4)
According to the methods used in this article, Clojure should be close to Java:
github.com/Taqmuraz/speed-table/is...
On my machine Clojure test takes 373ms. If I do benchmarking wrong, let me know
After the garbage collector moves or disposes of objects, there may be "holes" in the allocated memory unless it performs "defragmentation" afterward. So, newly created objects do not necessarily mean simply "moving the pointer to the right."
Additionally, the garbage collector is notoriously known for causing the entire system to wait while it completes its tasks. It runs unexpectedly and can take significant amount of time in large projects, so you literally cannot build "real-time" systems on GC languages
Benchmarks demonstrate opposite.
In Java or Go GC allocates memory just moving the pointer after filling by zeros. As I said in the article, GC moves all alive objects to the end of the last generation heap peak. So, there are no gaps between objects.