Memory allocation in D

[Sean Kelly]

Marius Muja wrote:
> Hi,
> 
> I'm working on a program that is computationally and memory intensive. I 
> have noticed that if I allocate the memory using D's new operator I can 
> allocate only about half as much memory compared to C's malloc (and also 
> the memory allocation is much slower). Does anybody know why D's memory 
> allocator uses twice as much memory compared to malloc?
> 
> I used the following test program to test this:
> 
> version (Tango){
> import tango.core.Memory;
> import tango.stdc.stdlib;
> import tango.io.Stdout;
> }
> else
> {
> import std.stdio;
> import std.c.stdlib;
> }
> 
> const int ALLOC_SIZE = 1024*1024;
> 
> void test_d_alloc()
> {
>     float[] a = new float[ALLOC_SIZE];
> }

With D as it is now, this will call GC.malloc(2049), which will allocate 
one page of memory per call.  The D runtime adds 1 to every allocation 
with "new", according to Walter, this was to prevent certain confusing 
errors.  It will allocate a page per call because the D GC allocates 
from pools devoted to fixed-size blocks which grow in powers of two up 
to 4096 bytes (one page on most systems).  Allocations beyond 4096 bytes 
are "big" allocations, and a multiple of 4096 bytes will be allocated to 
fit the requested size.  For these allocations, when the memory is 
released I believe the pages go back into a free pool.  I'm not sure 
offhand if empty pages used for smaller blocks do or not.

You might want to try calling GC.malloc() instead and compare results.


Sean