One task's record for the exchange profiler. More...

#include <exchangeoperator.h>

Public Member Functions
void	observe_fetch_tier (const int tier)
	keep the worst source, since that is the one that set the task's wait

void	reset ()

Public Attributes
double	apply_wall = 0.0

long	col_begin = 0

long	col_end = 0

double	compute_cpu = 0.0

double	compute_wall = 0.0

bool	diagonal = false

long	k = 0

double	mul1_wall = 0.0

double	mul2_wall = 0.0

int	operand_source = -1
	worst of its fetches: 0 resident, 1 ahead, 2 cold

double	peak_rss_gb = 0.0

long	row_begin = 0

long	row_end = 0

unsigned long	subworld_id = 0

int	subworld_nrank = 0

long	task_id = -1

double	thresh = 0.0
	which protocol tier this task ran in

double	truncate_wall = 0.0

long	universe_rank = 0
	keys the output file: one per process

double	wait_for_data_wall = 0.0
	task entry until its operands are in hand

bool	waited = false
	a cold fetch happened, so this task paid latency

double	wall_end = 0.0

double	wall_start = 0.0

Detailed Description

One task's record for the exchange profiler.

Written only when MAD_EXCH_TASK_PROFILE is set. What it adds over the aggregate counters is attribution: the counters say how many batches arrived from where, this says which task waited and for how long, so a straggler can be identified rather than inferred.

It deliberately does not carry a per-stage breakdown of the compute (multiply / apply / multiply). That would mean timing calls inside the numerical kernels, and the aggregate split already exists in the operator's own timers.

Member Function Documentation

◆ observe_fetch_tier()

void madness::ExchTaskProfile::observe_fetch_tier ( const int tier )

inline

keep the worst source, since that is the one that set the task's wait

References operand_source, and waited.

Referenced by madness::Exchange< T, NDIM >::ExchangeImpl< T, NDIM >::MacroTaskExchangeSimple::fetch_batch().

◆ reset()

void madness::ExchTaskProfile::reset ( )

inline

Referenced by madness::Exchange< T, NDIM >::ExchangeImpl< T, NDIM >::MacroTaskExchangeSimple::operator()().

Member Data Documentation

◆ apply_wall

double madness::ExchTaskProfile::apply_wall = 0.0

Referenced by madness::Exchange< T, NDIM >::ExchangeImpl< T, NDIM >::MacroTaskExchangeSimple::compute_diagonal_batch_in_symmetric_matrix().

◆ col_begin

long madness::ExchTaskProfile::col_begin = 0

Referenced by madness::Exchange< T, NDIM >::ExchangeImpl< T, NDIM >::MacroTaskExchangeSimple::operator()().

◆ col_end

long madness::ExchTaskProfile::col_end = 0

Referenced by madness::Exchange< T, NDIM >::ExchangeImpl< T, NDIM >::MacroTaskExchangeSimple::operator()().

◆ compute_cpu

double madness::ExchTaskProfile::compute_cpu = 0.0

Referenced by madness::Exchange< T, NDIM >::ExchangeImpl< T, NDIM >::MacroTaskExchangeSimple::operator()().

◆ compute_wall

double madness::ExchTaskProfile::compute_wall = 0.0

Referenced by madness::Exchange< T, NDIM >::ExchangeImpl< T, NDIM >::MacroTaskExchangeSimple::operator()().

◆ diagonal

bool madness::ExchTaskProfile::diagonal = false

Referenced by madness::Exchange< T, NDIM >::ExchangeImpl< T, NDIM >::MacroTaskExchangeSimple::operator()().

◆ k

long madness::ExchTaskProfile::k = 0

Referenced by madness::Exchange< T, NDIM >::ExchangeImpl< T, NDIM >::MacroTaskExchangeSimple::operator()().

◆ mul1_wall

double madness::ExchTaskProfile::mul1_wall = 0.0

wall inside each stage of the tile loop, accumulated over its rows. Honest without adding any fence: every stage below runs with fence=true, so each completes before the next is timed. What they do not cover – building the per-row update vector, the compresses and the accumulating gaxpys – shows up as the emitted other residual.

Referenced by madness::Exchange< T, NDIM >::ExchangeImpl< T, NDIM >::MacroTaskExchangeSimple::compute_diagonal_batch_in_symmetric_matrix().