Hi All,
I've fixed the MPI problems created by iwrap2. the code now passes all
tests, serial and parallel, withe the exception of pcsa, pb_ivcap1 and
pb_ivcap5, which seem to be unrelated to that problem.
I deeply apologize to everyone for any trouble this may have created.
Gustavo.
> On Oct 7, 2008, at 7:10 AM, David A. Case wrote:
>
>> (Mainly for Gustavo, but also everyone:)
>>
>> There appears to be serious problems with the recent iwrap2() stuff in
>> amber11: in sander.f, the allocate_molecule() routine only seems to be
>> called
>> by master nodes (which seems wrong), and is deallocated by all nodes. The
>> code then tries to broadcast the atom_to_resid_map to nodes where this
>> space
>> has never been allocated.
>>
>> The CVS logs even state "not ready for parallel" and "needs to be tested".
>>
>> We need to take more care with the CVS tree. Please remember that lots of
>> people depend on this tree mostly working. Don't commit half-finished
>> work;
>> don't commit stuff that won't pass the test cases.
>>
>> Also, do we really need an iwrap2() funtion and all the complexity it
>> entails?
>> I would prefer to remove all wrapping from sander (if possible), not to
>> make
>> it yet more complex. I know it's a losing battle to keep sander as simple
>> as
>> we can, but it's still should be a design criterion.
>>
>> ...thanks...dac
Received on Wed Oct 08 2008 - 05:10:44 PDT
