Calculator groups over MPI

This page explains how the MPI-parallel engines share one force call across the ranks of a calculator. The build option, the loader of librgpot_mpi and its C ABI are in MPI. The C++ functions (bindCalculators, shareFromCalculator, calculatorAgree and the rest) are in C++ API.

Collective force calls

CPMDPot and CP2KPot report PotCaps::groupCollective: every rank of a calculator enters each of its force calls together, and the call’s collectives stay on that calculator’s communicator. A host may hand calculators an uneven batch, with some calculators making no call at all, and then share results with shareFromCalculator. A failure inside the call prints on every rank of the failing calculator, then MPI_Abort on MPI_COMM_WORLD ends the run.

The result cache

For such a potential the result cache decides hit or miss jointly, through calculatorAgree (a logical AND over the calculator’s communicator). A geometry that any rank of the calculator misses is computed on all of its ranks, so no rank waits alone inside the call. Give each rank its own cache directory: RocksDB locks a directory to one process.

CP2K’s communicator

CP2KPot::bindCalculators splits the world as CPMDPot::bindCalculators does and hands the shim the calculator communicator through cp2kc_adopt_calculator_comm. It returns -1 without splitting when the shim does not load or lacks that symbol. Each force environment runs on that communicator, or on MPI_COMM_WORLD when no split exists. When MPI is initialised before the first session, the shim calls cp2k_init_without_mpi; otherwise cp2k_init starts MPI and the shim finalises CP2K and MPI at exit.

The CP2K tutorial is Evaluate diamond silicon with CP2K.