[Openmcl-devel] Fine-grained locking

[Daniel Herring]

On Tue, 17 Mar 2009, Ian Eslick wrote:

> Clozure is my performance target and I would like some input on the
> relative costs of two approaches to concurrency: fine-grained locking
> and copy-on-write.

I don't have the code/numbers handy, but I wrote a simple C test that 
timed various synchronization techniques.  If single async/plain memory 
write had a cost of 1, then a single uncontested CAS had a cost of ~10 and 
an uncontested pthread lock/unlock had a cost of ~100.

For many data structures, copy-on-write is a clear win since you don't 
need to copy the data, just a pointer/reference.

As far as locking patterns go,
- If you have multiple locks, design a full ordering for their use so you
never have two threads lock them in different orders.
- Lock the smallest sections possible.
- Lock as few times as possible.

Later,
Daniel