1 comment

[ 3.2 ms ] story [ 9.6 ms ] thread
Maybe you can have an implementation of shared-memory model that is "predictable" (transactional). But I failed to see how that solves the synchronization problem shared-memory model exposed to at large-scale system.

On the other hand, GPGPU community has an interesting idea to this problem. The idea is to have a memory hierarchy and at the bottom of the level, you have a homogeneous continuous memory space which is painfully slow and have strong consistency. When you walk up the hierarchy, you will have faster, smaller, local memory that have weaker consistency.

Disclaimer: I haven't got time around to read the paper throughly, it is just my two cents.