Simon Marlow
On 17/08/2010 06:09, Gregory Collins wrote:
We could provide a compare-and-swap on IORefs, indeed I've been planning to try that so we could move the implementation of atomicModifyIORef into user-space, so to speak.
For hash tables it would be nice to have a CAS indexing primitive on MutableArray#/MutableByteArray#/etc, also.
Under load, pure data structures behind MVars don't perform well because of lock contention, and in my experience atomicModifyIORef is marginally better but still suffers from contention issues (probably related to thunk blackholing).
This is the specific case that we addressed in the GHC runtime recently, and you should find that 6.14 will perform a lot better than 6.12 with pure data structures in mutable containers, especially with atomicModifyIORef.
In the meantime you'll probably get better performance using STM. Whether you should perform the update strictly inside the transaction or not is a matter for experimentation, it probably depends on the data structure.
In the specific benchmark in question STM was slightly slower than MVars
(but that's pretty impressive!).
G
--
Gregory Collins