I'd be interested to know whether you, or others, find this custom memory allocation interface satisfactory. After having used it I think that I quite like this approach because it allows the developer to use application-specific knowledge about the number of concurrent asynchronous "chains" when customising memory allocation.
This custom memory allocation implementation required no changes to the existing asio public interface or overall design.
Chris, I'm trying to catch up here after hacking on the reactor implementation in a corner for a few days. Could you point me to a post that describes the custom memory allocation interface? One thing I've done is written a pooled_list implementation and parametrized the list implementation used by your hash_map. This eliminates a bunch of hits to the global allocator, at the cost of a pre-allocated fixed hash_map size. Considering that the hash_map associates FDs with handlers, I think this is a reasonable trade off. christopher