I tested a few of the approaches suggested here.
I allocated a huge set of data (200GB) in one container/pointer:
Compiler/OS:
g++ (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
Settings: (c++-17, -O3 optimizations)
g++ --std=c++17 -O3
I timed the total program runtime with linux-time
1.) std::vector:
#include <vector>
int main(){
constexpr size_t size = 1024lu*1024lu*1024lu*25lu;//25B elements = 200GB
std::vector<size_t> vec(size);
}
real 0m36.246s
user 0m4.549s
sys 0m31.604s
That is 36 seconds.
2.) std::vector with boost::noinit_adaptor
#include <vector>
#include <boost/core/noinit_adaptor.hpp>
int main(){
constexpr size_t size = 1024lu*1024lu*1024lu*25lu;//25B elements = 200GB
std::vector<size_t,boost::noinit_adaptor<std::allocator<size_t>>> vec(size);
}
real 0m0.002s
user 0m0.001s
sys 0m0.000s
So this solves the problem. Just allocating without initializing costs basically nothing (at least for large arrays).
3.) std::unique_ptr<T[]>:
#include <memory>
int main(){
constexpr size_t size = 1024lu*1024lu*1024lu*25lu;//25B elements = 200GB
auto data = std::unique_ptr<size_t[]>(new size_t[size]);
}
real 0m0.002s
user 0m0.002s
sys 0m0.000s
So basically the same performance as 2.), but does not require boost.
I also tested simple new/delete and malloc/free with the same performance as 2.) and 3.).
So the default-construction can have a huge performance penalty if you deal with large data sets.
In practice you want to actually initialize the allocated data afterwards.
However, some of the performance penalty still remains, especially if the later initialization is performed in parallel.
E.g., I initialize a huge vector with a set of (pseudo)random numbers:
(now I use fopenmp for parallelization on a 24 core AMD Threadripper 3960X)
g++ --std=c++17-fopenmp -O3
1.) std::vector:
#include <vector>
#include <random>
int main(){
constexpr size_t size = 1024lu*1024lu*1024lu*25lu;//25B elements = 200GB
std::vector<size_t> vec(size);
#pragma omp parallel
{
std::minstd_rand0 gen(42);
#pragma omp for schedule(static)
for (size_t i = 0; i < size; ++i) vec[i] = gen();
}
}
real 0m41.958s
user 4m37.495s
sys 0m31.348s
That is 42s, only 6s more than the default initialization.
The problem is, that the initialization of std::vector is sequential.
2.) std::vector with boost::noinit_adaptor:
#include <vector>
#include <random>
#include <boost/core/noinit_adaptor.hpp>
int main(){
constexpr size_t size = 1024lu*1024lu*1024lu*25lu;//25B elements = 200GB
std::vector<size_t,boost::noinit_adaptor<std::allocator<size_t>>> vec(size);
#pragma omp parallel
{
std::minstd_rand0 gen(42);
#pragma omp for schedule(static)
for (size_t i = 0; i < size; ++i) vec[i] = gen();
}
}
real 0m10.508s
user 1m37.665s
sys 3m14.951s
So even with the random-initialization, the code is 4 times faster because we can skip the sequential initialization of std::vector.
So if you deal with huge data sets and plan to initialize them afterwards in parallel, you should avoid using the default std::vector.
std::allocatoronly has value and zero initialization, but has no method for default initialization. – Mllly