Atom deletion is very slow

[Tom Goddard]

When deleting atoms from a molecule each atom delete causes all DestructionObserver objects to be called. This leads to extremely slow order N*N speed when deleting the atom from Python arrays of atoms because for each deleted atom, the entire array has to be scanned to delete that atom. There is a Python array of all the atom pointers for the graphics.

What would help is having an AtomicStructure delete_atoms() method that can take a whole list of atoms and produce a single DestructionObserver call.

I don't have measured times for deleting many atoms because other bugs are causing crashes when deleting atoms.

[Eric Pettersen]

Implemented delete_atoms(). Still need to properly delete pseudobonds.

[Tom Goddard]

Deleting chain 0 (65,000 atoms) from pdb 1jj2 (100,000) takes 95 seconds with the new delete_atoms() call. Investigation shows the destruction observers are only called once on 131,000 items (atoms and bonds) and that processing is handled instantly (fraction of a second). All the time is in the actual C++ delete_atoms(). That code removes the atoms and bonds from a vector one at a time. It searches linearly through the vector then erases the element causing all the items in the vector to be shifted down in memory. This is requiring on the order of 2 * 65,000 * 100,000 = 13 billion operations for search and delete, so it is no surprise that it is slow. It is probably necessary to use a std::set of deleted atoms, make a set of deleted bonds, then do one pass on the relevant vectors to eliminate all the deleted atoms and bonds, by building new vectors of the surviving atoms and bonds.