gmx::PmeLoadBalancing::Impl Class Reference

Description

Impl class for PmeLoadBalancing.

The states the algorithm can reside in:

• isActive_ == false: inactive, after tuning has finished

• isActive_ == true:

  • isInBalancingPhase == false: only timing and checking for imbalance, sufficient imbalance will trigger the balancing phase
  • isInBalancingPhase == true:
    • stage == 0: generate and step coarsely through a range of grid to get an initial grid range and rough timings
    • stage > 0: time grids within a sub-range of interest, this sub-range can be extended when the timings change a lot from the initial ones

  Will, permanently, move to isActive_ = false when no imbalance is observed over many steps or when the balancing has finished.

Public Member Functions
	Impl (gmx_domdec_t *dd, const MDLogger &mdlog, const t_inputrec &ir, const matrix box, const interaction_const_t &ic, const nonbonded_verlet_t &nbv, gmx_pme_t *pmedata, const SimulationWorkload &simulationWork)
bool	isActive () const
bool	isPrintingLoad () const
int	setupEnd () const
	Returns one index past the last setup that will be covered during balancing.
bool	increaseCutoff ()
	Attempts to increase the cutoff, returns true when a setup has been added to the list.
void	switchToStage1 ()
	Switch load balancing to stage 1. More...
void	balance (FILE *fp_err, const matrix box, ArrayRef< const RVec > x, double cycles, interaction_const_t *ic, nonbonded_verlet_t *nbv, gmx_pme_t **pmedata, int64_t step)
	Process the timings and try to adjust the PME grid and Coulomb cut-off. More...
void	addTwoStages (bool dlbWasUnlocked)
	Prepare for another round of PME load balancing. More...
void	addCycles (FILE *fp_err, t_forcerec *fr, const matrix box, ArrayRef< const RVec > x, const gmx_wallcycle *wcycle, int64_t step, int64_t step_rel)
void	printSettings () const

Member Function Documentation

void gmx::PmeLoadBalancing::Impl::addTwoStages

(

bool

dlbWasUnlocked

)

Prepare for another round of PME load balancing.

Parameters

[in]

dlbWasUnlocked

Pass true when DLB was locked and is now unlocked

If the conditions (e.g. DLB off/on, CPU/GPU throttling etc.) changed, the PP/PME balance might change and re-balancing can improve performance. This function adds 2 stages and adjusts the considered setup range.

void gmx::PmeLoadBalancing::Impl::balance	(	FILE *	fp_err,
		const matrix	box,
		ArrayRef< const RVec >	x,
		double	cycles,
		interaction_const_t *	ic,
		nonbonded_verlet_t *	nbv,
		gmx_pme_t **	pmedata,
		int64_t	step
	)

Process the timings and try to adjust the PME grid and Coulomb cut-off.

The adjustment is done to generate a different non-bonded PP and PME load. With separate PME ranks (PP and PME on different processes) or with a GPU (PP on GPU, PME on CPU), PP and PME run on different resources and changing the load will affect the load balance and performance. The total time for a set of integration steps is monitored and a range of grid/cut-off setups is scanned. After calling pme_load_balance many times and acquiring enough statistics, the best performing setup is chosen. Here we try to take into account fluctuations and changes due to external factors as well as DD load balancing.

void gmx::PmeLoadBalancing::Impl::switchToStage1

(

)

Switch load balancing to stage 1.

In this stage, only sufficiently fast setups are run again.

---

The documentation for this class was generated from the following file:

• src/gromacs/ewald/pme_load_balancing.cpp