Thanks for the patch. I deplore the lost of the hash table in the attribute management, as the potential of transforming all attributes operation to a linear complexity is not very appealing.
As you already took the decision C, it means that at the communicator destruction stage the hash table is not relevant anymore. Thus, I would have converted the hash table to an ordered list (ordered by the creation index, a global entity atomically updated every time an attribute is created), and proceed to destroy the attributed in the desired order. Thus instead of having a linear operation for every operation on attributes, we only have a single linear operation per communicator (and this during the destruction stage).
On Jan 16, 2013, at 16:37 , KAWASHIMA Takahiro <rivis.kawashima_at_[hidden]> wrote:
> I've implemented ticket #3123 "MPI-2.2: Ordering of attribution deletion
> callbacks on MPI_COMM_SELF".
> As this ticket says, attributes had been stored in unordered hash.
> So I've replaced opal_hash_table_t with opal_list_t and made necessary
> modifications for it. And I've also fixed some multi-threaded concurrent
> (get|set|delete)_attr call issues.
> By this modification, following behavior changes are introduced.
> (A) MPI_(Comm|Type|Win)_(get|set|delete)_attr function may be slower
> for MPI objects that has many attributes attached.
> (B) When the user-defined delete callback function is called, the
> attribute is already removed from the list. In other words,
> if MPI_(Comm|Type|Win)_get_attr is called by the user-defined
> delete callback function for the same attribute key, it returns
> flag = false.
> (C) Even if the user-defined delete callback function returns non-
> MPI_SUCCESS value, the attribute is not reverted to the list.
> (A) is due to a sequential list search instead of a hash. See find_value
> function for its implementation.
> (B) and (C) are due to an atomic deletion of the attribute to allow
> multi-threaded concurrent (get|set|delete)_attr call in MPI_THREAD_MULTIPLE.
> See ompi_attr_delete function for its implementation. I think this does
> not matter because MPI standard doesn't specify behavior in such cases.
> The patch for Open MPI trunk is attached. If you like it, take in
> this patch.
> Though I'm a employee of a company, this is my independent and private
> work at my home. No intellectual property from my company. If needed,
> I'll sign to Individual Contributor License Agreement.
> KAWASHIMA Takahiro
> devel mailing list