Comparing performance with QMCPACK

The key kernels from QMCPACK are present in miniQMC, but the relative importance of each kernel may not be retained.

The inverse determinant update is an N^3 operation, and increasingly dominates the run time as the system size increases.

The acceptance ratio option (-r) affects the number of inverse updates performed. The default setting is 0.5, which corresponds the acceptance ratio of a VMC run. To more accurately reproduce a DMC run, use an acceptance ratio 0.99 (-r 0.99)

To more closely reproduce the number of calls from the DMC NiO test cases, use "-n 25 -r 0.999".

The non-local pseudopotential evaluation in QMCPACK is also part of the miniapp, but is not part of the core particle movement loop.

Kernel	QMCPACK timer name	MiniQMC timer name
Total (core particle loop)	DMCUpdatePbyP::movePbyP	Diffusion
Inverse matrix update	DiracDeterminantBase::update	Update/Determinant
Single Particle Orbitals	DiracDeterminantBase::spovgl	New Gradient/Single-Particle Orbitals
Two body Jastrow	WaveFunction::J2_bspline_accept_reject	Update/TwoBodyJastrow
...	WaveFunction::J2_bspline_VGL	New Gradient/TwoBodyJastrow
Distance tables	ParticleSet::makeMove	Diffusion/Make move
...	ParticleSet::setActive	Diffusion/Set active
...	ParticleSet::acceptMove	Diffusion/Accept move
Non-local pseudopotential (not part of core loop)	Hamiltonian::NonLocalECP	Pseudopotential

Comparing times

Time fraction for MiniQMC vs. QMCPACK. Run on dual-socket Skylake 8180. First grouping of bars (lighter colors) is miniQMC. Second grouping of bars (darker colors) is QMCPACK. Each bar in the grouping is a different thread count (1, 14, 28, and 56).

Absolute times using "-n 25 -r 0.999". These runs use 14 threads.