Re: [Development] Proposal for "container-oriented deterministic memory manager"

[Phil Bouchard]

On 01/07/2017 12:11 PM, Phil Bouchard wrote:
On 01/06/2017 10:37 PM, Phil Bouchard wrote:
On 01/06/2017 08:05 PM, Phil Bouchard wrote:
On 01/06/2017 07:17 PM, Phil Bouchard wrote:

Just to conclude I did try the attached benchmark and I get the
following on a x86_64 @ 2.40 GHz:

0: 61331143.40263957 allocations / second
1: 63644162.93924019 allocations / second
2: 177628727.5388474 allocations / second
3: 179850939.5413082 allocations / second
1 / 0: 103.7713621632905% boost
2 / 1: 101.2510431354494% boost

So the fast_pool_allocator is already pretty fast and I can only get a
1% speed boost by allocating big memory blocks.  So it doesn't look
like
there is any way to make the fast_pool_allocator any faster.

Correction: I get a 11% speed boost with the -O3 flag:

0: 60387275.67636769 allocations / second
1: 63133704.55951615 allocations / second
2: 169169529.8609596 allocations / second
3: 188535531.4062488 allocations / second
1 / 0: 104.5480258090584% boost
2 / 1: 111.447688931456% boost

So I am not sure if this is worth the trouble.

Did you know the optimized boost::fast_pool_allocator is 305% times
faster than the regular one?

0: 61407861.18875794 allocations / second
1: 62565419.96725326 allocations / second
2: 175311265.1512761 allocations / second
3: 187319938.708916 allocations / second
1 / 0: 101.8850335381934% boost
2 / 1: 106.8499155187076% boost
3 / 0: 305.0422781101666% boost

Is Qt using a similar pool allocator?

This means the optimized boost::fast_allocator is 470% times faster than
the system "operator new":

0: 37334468.30116969 allocations / second (operator new)
1: 58554420.18534816 allocations / second (boost::pool_allocator)
2: 60072218.82146717 allocations / second (boost::pool_allocator)
3: 164895686.9884111 allocations / second (boost::fast_pool_allocator)
4: 175697519.1510296 allocations / second (boost::fast_pool_allocator)
2 / 1: 102.5921162421463% boost
4 / 3: 106.550706304027% boost
4 / 0: 470.6040480708413% boost

I just tried the "jemalloc" library and I get a 2576228000% speed boost 
compared to operator new:

0: 38816440.15979952 allocations / second
1: 62188319.91413037 allocations / second
2: 63819182.38692677 allocations / second
3: 174727730.5139267 allocations / second
4: 186025745.9632413 allocations / second
5: 1000000000000000 allocations / second
5 / 0: 2576228000% boost
5 / 1: 1608019000% boost
#include <limits>
#include <iomanip>
#include <boost/timer.hpp>
#include <boost/pool/pool_alloc.hpp>
#define JEMALLOC_NO_DEMANGLE
#include "jemalloc/jemalloc.h"

using namespace std;
using namespace boost;

int main()
{
    double speed[6];
    
    static long const n = 100000000;
    
    {
        timer t;
        
        for (int i = 0; i < n; ++ i)
            ::operator new (sizeof(int));
        
        speed[0] = n / t.elapsed();
    }

    {
        pool_allocator<int, default_user_allocator_new_delete, details::pool::default_mutex> p;
        timer t;
        
        for (int i = 0; i < n; ++ i)
            p.allocate(1);
        
        speed[1] = n / t.elapsed();
    }

    {
        pool_allocator<int, default_user_allocator_new_delete, details::pool::default_mutex, n> p;
        timer t;
        
        for (int i = 0; i < n; ++ i)
            p.allocate(1);
        
        speed[2] = n / t.elapsed();
    }

    {
        fast_pool_allocator<int, default_user_allocator_new_delete, details::pool::default_mutex> p;
        timer t;
        
        for (int i = 0; i < n; ++ i)
            p.allocate(1);
        
        speed[3] = n / t.elapsed();
    }

    {
        fast_pool_allocator<int, default_user_allocator_new_delete, details::pool::default_mutex, n> p;
        timer t;
        
        for (int i = 0; i < n; ++ i)
            p.allocate(1);
        
        speed[4] = n / t.elapsed();
    }
    
    {
        timer t;
        
        for (int i = 0; i < n * 10; ++ i)
            je_malloc(sizeof(int));
        
        speed[5] = n * 10 / t.elapsed();
    }

    cout << setprecision(numeric_limits<double>::digits10 + 1);
    
    for (int i = 0; i < 6; ++ i)
        cout << i << ": " << speed[i] << " allocations / second" << endl;

    //for (int i = 1; i < 4; i += 2)
    //    cout << i + 1 << " / " << i << ": " << speed[i + 1] / speed[i] * 100 << "% boost" << endl;

    cout << 5 << " / " << 0 << ": " << speed[5] / speed[0] * 100 << "% boost" << endl;
    cout << 5 << " / " << 1 << ": " << speed[5] / speed[1] * 100 << "% boost" << endl;

    return 0;
}
_______________________________________________
Development mailing list
Development@qt-project.org
http://lists.qt-project.org/mailman/listinfo/development