Memory and Performance Overhead of Smart Pointers

C++11 offers four different smart pointers. On two of them, I will have a closer look in this post regarding memory and performance overhead. My first candidate std::unique_ptr takes exclusively care of the lifetime of one resource; std::shared_ptr shares the ownership of a resource with other std::shared_ptr's. I will state the result of my tests before I show you the raw numbers: There are only a few reasons in modern C++ justifying the memory management with new and delete.

Why? Here are the numbers.

Memory overhead

std::unique_ptr

std::unique_ptr needs by default no additional memory. That means, std::unique_ptr is as big as its underlying raw pointer. But what does by default mean? You can parametrize a std::unique_ptr with a special deleter function. If this deleter function has been stated, you will have an enriched std::unique_ptr and you will pay for it. As I mentioned, that is the special use case.

In opposite to the std::unique_ptr the std::shared_ptr has a little overhead.

std::shared_ptr

std::shared_ptr's share a resource. They internally use a reference counter. That means, if a std::shared_ptr is copied, the reference counter will be increased. The reference count will be decreased if the std::shared_ptr goes out of scope. Therefore, the std::shared_ptr needs additional memory for the reference counter. (To be precise, there is an additional reference counter for the std::weak_ptr). That's the overhead a std::shared_ptr has in opposite to a raw pointer.

Performance overhead

The story of the performance is a little bit more involved. Therefore, I let the numbers speak for themself. A simple performance test should give an idea of the overall performance.

The performance test

I allocate and deallocate in the test 100'000'000 times memory. Of course, I'm interested in how long it takes.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

// all.cpp

#include <chrono>
#include <iostream>

static const long long numInt= 100000000;

int main(){

  auto start = std::chrono::system_clock::now();

  for ( long long i=0 ; i < numInt; ++i){
    int* tmp(new int(i));
    delete tmp;
    // std::shared_ptr<int> tmp(new int(i));
    // std::shared_ptr<int> tmp(std::make_shared<int>(i));
    // std::unique_ptr<int> tmp(new int(i));
    // std::unique_ptr<int> tmp(std::make_unique<int>(i));
  }

  std::chrono::duration<double> dur= std::chrono::system_clock::now() - start;
  std::cout << "time native: " << dur.count() << " seconds" << std::endl;

}

I compare in my test the explicit calls of new and delete (line 13 and 14) with the usage of  std::shared_ptr (line 15), std::make_shared (line 16), std::unique_ptr (line 17), and std::make_unique (line 18). Already in this small program, the handling of the smart pointer (line 15 - 18) is a lot simpler because the smart pointer automatically releases its dynamically created int variable if it goes out of scope.  

The two functions std::make_shared (line 16) and std::make_unique(line 18) are quite handy. They create the smart pointer respectively. In particular, std::make_shared is very interesting. There are more memory allocations necessary for the creation of a std::shared_ptr. Memory is necessary for the managed resource and the reference counters. std::make_shared make one memory allocation out of them. The performance benefits. std::make_unique is available since C++14; the other smart pointer functionality since C++11.

I use a GCC 4.9.2 and a cl.exe for my performance tests. The cl.exe is part of the Microsoft Visual Studio 2015. Although cl.exe officially supports only C++11, the helper function std::make_unique is already available. Therefore, I can run my performance test with maximum and without optimization. I have to admit that my Windows PC is not as powerful as my Linux PC. Therefore, I'm interested in the comparison between the raw memory allocation and the smart pointers. I'm not comparing Windows and Linux.

Here a the raw numbers.

The raw numbers

For simplicity reasons, I will not show the screenshots of the programs and present you only the table holding the numbers.

I want to draw a few interesting conclusions from the table.

1. Optimization matters. In the case of std::make_shared_ptr the program with maximum optimizations is almost 10 times faster. But these observations holds also for the other smart pointers. The optimized program is about 2 to 3 times faster. Interestingly, this observation will not hold for new and delete.
2. std::unique_ptr, std::make_unique, and with small deviations std::make_shared are in the same performance range as new and delete.
3. You should not use std::shared_ptr and std::make_shared without optimization. std::shared_ptr is even with optimization about two times slower than new and delete.

My conclusion

• std::unique_ptr has no memory or performance overhead compared to the explicit usage of new and delete. That is very great because std::unique_ptr offers a great benefit by automatically managing the lifetime of its resource without any additional cost.
• My conclusion to std::shared_ptr is not so easy. Admittedly, the std::shared_ptr is about two times slower than new and delete. Even std::make_shared has a performance overhead of about 10%. But this calculation is based on the wrong assumptions because std::shared_ptr models shared ownership. That means, only the first std::shared_ptr has to shoulder the performance and memory overhead. The additional shared pointer shares the infrastructure of managing the underlying object. In particular, only one memory allocation for a std::shared_ptr is necessary. 

Now I can repeat myself.  There are only a few reasons in modern C++ justifying the memory management with new and delete.

What's next?

After this plea for the smart pointers, I will present in the next post the details about std::unique_ptr.