itaps-parallel A question about iMeshP interface functions
Onkar Sahni
osahni at scorec.rpi.edu
Thu Aug 14 09:29:37 CDT 2008
>>
>> These functions include:
>> iMeshP_getEntOwnerPart(), iMeshP_getEntOwnerPartArr(),
>> iMeshP_getNumCopies(), iMeshP_getCopyParts(),
>> iMeshP_getCopies(), iMeshP_getOwnerCopy().
>>
>>
>>
>
> I'm not sure what the issue is. Are you suggesting some alternate form of
> these functions? You began your initial message with a statement about
> functions for which the first two arguments are a partition handle and an
> entity handle. I assume you are not suggesting that we remove the
> partition
> handle argument as that would just increase the potential search space for
> these queries (naive implementations are still O(n) but N is larger if the
> search space is not constrained to a partition.)
>
> All of these functions are inherently expensive as they imply a search of
> at
> least the local mesh on a processor. There are, of course, speed vs
> memory
> trade offs possible in the implementation (the simple O(n) search you
> described in your original message certainly isn't the only possible
> implementation.)
Ting is not suggesting alternate form of these functions. I think the
question is related to error message/code, as per specification it says
"return an error code if an entity is not in the partition". How does an
implementation return this error if the entity passed is a remote one
(non-local to processor or off-proc.). Infact in the partition can mean
in any part (local and non-local).
These functions are not inherently expensive and we do not want to do
any search on the local mesh on a processor in these functions.
- Onkar
More information about the itaps-parallel
mailing list