Re: Really easy optimization with std.experimental.allocator

[Dicebot via Digitalmars-d-learn]

On Sunday, 18 September 2016 at 01:44:10 UTC, Ryan wrote:
I think it works because each time you call dispose it tells 
the GC to mark that memory as available, without the GC needing 
to do a collection sweep. This could be a really useful tip in 
the allocators section, as I see converting to IAllocator with 
the GC as the first step in testing optimizations with 
allocators.


A bit more than that - because you dispose and allocate same 
amount of memory, you effectively reuse same memory block in GC 
pools over and over again. It is hardly surprising that this is 
much faster than "honest" allocation of lot of memory across 
different pools (CPU memory cache access pattern alone will make 
a huge difference).

Really easy optimization with std.experimental.allocator

[Ryan via Digitalmars-d-learn]

I've been learning about allocators in D. Much easier than C++, 
and this little program shows a really easy optimization. Just 
use the IAllocator interface with the GC.

---
import std.datetime;
import std.experimental.allocator;
import std.experimental.allocator.mallocator;
import std.stdio;

enum sizes = [10,100, 1_000, 10_000];

// A function that creates several arrays on the GC heap with 
new, and assigns

// a single value, so it doesn't get compiled out
void GCNew()
{

  byte[] b1;

  foreach(sz; sizes)
  {
b1 = new byte[](sz);
b1[1] = 10;
assert(b1[1] == 10);
assert(b1[0] == 0);
  }
}

// Same as the function above, but with an IAllocator interface
void IAlloc(IAllocator alloc = theAllocator)
{
  byte[] b1;

  foreach(sz; sizes)
  {
b1 = alloc.makeArray!byte(sz);
b1[1] = 10;
assert(b1[1] == 10);
assert(b1[0] == 0);
alloc.dispose(b1);
  }
}

void main()
{
  enum iterations = 1_000;

  writefln("GCNew: %d", 
benchmark!GCNew(iterations)[0].usecs);


  writefln("IAlloc GC: %d", 
benchmark!IAlloc(iterations)[0].usecs);


  writefln("IAlloc malloc: %d", benchmark!(
()
{
  IAlloc(allocatorObject(Mallocator.instance));
})
  (iterations)[0].usecs);
}

---
Results on my iMac with rdmd -release -boundscheck=off -O 
allocTest.d

GCNew: 7467
IAlloc GC: 1657
IAlloc malloc: 1575

I think it works because each time you call dispose it tells the 
GC to mark that memory as available, without the GC needing to do 
a collection sweep. This could be a really useful tip in the 
allocators section, as I see converting to IAllocator with the GC 
as the first step in testing optimizations with allocators.