MPQC workshop 2006

Agenda

o coffee and bagels

o formulate agenda

- parallel infrastructure

- matrix package

- I/O

- CCA integral interfaces (derivatives, vectorization) + paper

- Fock builder

- LCC + F12

- future

parallel infrastructure

o Overview: MessageGrp, MemoryGrp, ThreadGrp. Reviewed the facilities.

o Several implementation of each are available. MPIMessageGrp, PthreadThreadGrp,

and MTMPI and ARMCI based MemoryGrp are mainly used (MTMPI support is lacking

across platforms, but recent MPICH2 supposedly is thread-safe across sockets).

o Abstraction is key because most packages are not portable, so should

be hidden behind an abstract interface.

o ActiveMessageGrp is desired for some operations.

o The infrastructure has been tested very heavily, never has to be

debugged unless the layer underneath is broken.

matrix package

o Repl, Local work

o new package (sparse, etc.) -- no good candidates, no funding

o LCC code uses custom impl based on AVLMap

I/O

o Used only in MP2-F12 code

o LIBPSIO was imported from Psi3 but not used anywhere (works?)

o FileGrp -- not finished (ProcFileGroup exists and may work) but would

be very useful (desired implementations: MPI-IO-based and distributed scratch)

CCA Integral Interfaces

o Long discussion about details of the interface. Redesigned IntegralDescr hierarchy to allow

arbitrary combination of integral types produced. CCA Integrals paper will need the interface

implementation.

o Brief review of the current draft of the paper.

Fock Builder

o Exists, but no symmetry. Should allow density fitting. Curt put it on libint2-branch.

Local correlation + F12

o Local MP2 paper is prepared.

o Local CC code exists and works. Some issues with release.

o Discussion of LMP2-F12: should not be difficult. Explicit correlation is probably

only important for strong pairs. Maximum storage is for A (part of H0-E0 which couples conventional

doubles and geminal doubles), which is same size as T2.

o Maximum extent of geminal factors in use today is very short, so should not affect the size of domains.

o Diffuse functions in RI basis should also pose no significant problems, at least

when one-RI-index-only approach is used, since RI functions are always multiplied by the orbital basis.

Future

o 2.4 to be released by the end of the year. New features: open-shell MP2-F12 (including app B and

perhaps C). CCA integral interfaces. LMP2? Most of current development of MP2-F12 and CCA is

on libint2-branch, which is synced with the main branch regularly.