core.atomic: atomicFence, atomicLoad, atomicStore

[John Colvin via Digitalmars-d-learn]

On Monday, 12 January 2015 at 13:37:19 UTC, ref2401 wrote:
> Thanks for the links.
>
> I have shared class instance. There are two threads which can 
> read/write fields of the class. As i understand i can declare 
> class as synchronized or i can read/write using 
> atomicLoad/atomicStore.
> What's the difference between these two approaches?
> In what circumstances should i consider synchronized classes or 
> using std.atomic?

synchronized uses locks, and as such is good for (rule-of-thumb) 
low contention situations, i.e. you don't have lots of threads 
all trying to access the same data at the same time (or few 
threads doing lots of repeated access). It has significant 
overhead as there is a cost to taking and releasing the lock and 
can be very slow under high-contention. Its advantage is that 
it's merely difficult to get right.

Using core.atomic and lock-free techniques in general is great 
for getting good performance in high-contention situations and/or 
reducing overhead in low-contention environments. However, it is 
very difficult to get right, very difficult to know if you have 
got it right and insanely difficult to debug when things go wrong.