The GC and performance, but not what you expect

[Adam Sakareassen via Digitalmars-d]

Hi,

Yes my little GC project is coming along.  It allocates much faster in 
multi threaded applications when all threads are competing for the lock. 
  D's current GC is very slow when there is contention for the GC-lock 
which is acquired for every malloc.  The other hidden lock usage is when 
a dynamic array increases in size.  A significant problem is that the 
thread holding the lock might be pre-empted by the OS, causing all 
threads to wait until it is running again.  My lock-free allocator fixes 
all this trouble.

(Note: On Windows the current GC uses EnterCriticalSection from 
kernel32.lib)

In single threaded applications, D's GC performs much better.  It still 
has the overhead of maintaining a free list while allocating. The free 
list is built when requesting pools from the OS.  This list is stored as 
a linked list using the fresh memory cells.  A cell is zerod when it is 
allocated.

My implementation is slightly faster at allocations because it doesn't 
need to zero the memory before before. It only maintains a such a free 
list for recycled memory.

I'll try to make some code available soon for people to try if they are 
interested.  There are still some jobs I need to do.  I'm just 
implementing the utility functions now (sizeOf, addrOf) etc.   Allocated 
memory is currently never returned to the OS, just recycled for new 
allocations.

Cheers
Adam








On 30/05/2014 6:29 PM, Rainer Schuetze via Digitalmars-d wrote:
>
>
> On 29.05.2014 12:09, Atila Neves wrote:
>> The GC is preventing me from beating Java, but not because of
>> collections. It's the locking it does to allocate instead! I
>> don't know about the rest of you but I definitely didn't see that
>> one coming.
>>
>
> A lock should not be more than a CAS operation that always succeeds in a
> single threaded application, but OS overhead might be considerable.
>
> Adam Sakareassen has recently announced a rewrite of the GC that has
> lock-free allocations:
> http://forum.dlang.org/thread/mailman.655.1399956110.2907.digitalmars-d@puremagic.com
>
>
> Unfortunately, it isn't available yet...
>