Collective benchmarks

Classes
struct	MPIB_coll_container
Typedefs
typedef int(*	MPIB_measure_coll )(MPIB_coll_container *container, MPI_Comm comm, int root, int M, MPIB_precision precision, MPIB_result *result)
typedef void(*	MPIB_measure_coll_msgset )(MPIB_coll_container *container, MPI_Comm comm, int root, MPIB_msgset msgset, MPIB_precision precision, int *count, MPIB_result **results)
Functions
int	MPIB_measure_max (MPIB_coll_container *container, MPI_Comm comm, int root, int M, MPIB_precision precision, MPIB_result *result)
void	MPIB_root_timer_init (MPI_Comm comm, int reps)
int	MPIB_measure_root (MPIB_coll_container *container, MPI_Comm comm, int root, int M, MPIB_precision precision, MPIB_result *result)
void	MPIB_global_timer_init (MPI_Comm comm, int parallel, int reps)
int	MPIB_measure_global (MPIB_coll_container *container, MPI_Comm comm, int root, int M, MPIB_precision precision, MPIB_result *result)

---

Detailed Description

This module provides collective benchmarks based on root, max and global timing methods (see ).

---

Typedef Documentation

typedef int(* MPIB_measure_coll)(MPIB_coll_container *container, MPI_Comm comm, int root, int M, MPIB_precision precision, MPIB_result *result)

Collective benchmark. Measures the execution time of collective operation for a given message size.

Parameters:

	container	communication operaion container
	comm	communicator
	root	root process
	M	message size
	precision	measurement precision
	result	measurement result

typedef void(* MPIB_measure_coll_msgset)(MPIB_coll_container *container, MPI_Comm comm, int root, MPIB_msgset msgset, MPIB_precision precision, int *count, MPIB_result **results)

Collective benchmark for multiple message sizes. Measures the execution time of collective operation for different message sizes.

Parameters:

	container	communication operaion container
	comm	communicator
	root	root process
	msgset	message sizes
	precision	measurement precision
	count	the number of measurements performed (significant only at the root processor)
	results	array of measurement results (significant only at the root processor, allocated by this function and must be deallocated by user)

---

Function Documentation

int MPIB_measure_max	(	MPIB_coll_container *	container,
		MPI_Comm	comm,
		int	root,
		int	M,
		MPIB_precision	precision,
		MPIB_result *	result
	)

Measures the execution time of collective operation at all processes and finds a maximum. In the loop over repetitions:

• Synchronizes the processes by double barrier.
• Measures the execution time of collective operation at all processes.
• Finds maximum by allreduce.
• Performs statistical analysis.

void MPIB_root_timer_init	(	MPI_Comm	comm,
		int	reps
	)

Measures the average execution time of barrier at all processes in the communicator and sets up an internal global variable, which is used by MPIB_measure_root. Called by MPIB_measure_root. Can be called directly before measurements.

Parameters:

	comm	communicator
	reps	number of repetitions (not precision - we assume no pipeline effect with barrier)

int MPIB_measure_root	(	MPIB_coll_container *	container,
		MPI_Comm	comm,
		int	root,
		int	M,
		MPIB_precision	precision,
		MPIB_result *	result
	)

Measures the execution time of collective operation at the root process. Reuses already obtained barrier time if the previous initialization was performed over the same communicator. Otherwise, initializes the root timer by calling MPIB_root_timer_init.
In the loop over repetitions:

• Synchronizes the processes by double barrier.
• Measures the execution time of collective operation and barrier confirmation at the root process.
• Subtracts the execution time of barrier.
• Performs statistical analysis at root.

Broadcasts the result.

void MPIB_global_timer_init	(	MPI_Comm	comm,
		int	parallel,
		int	reps
	)

Measures the offsets between local clocks of of all processes in the communicator and sets up an internal global variable, which is used by MPIB_measure_global. If MPI_WTIME_IS_GLOBAL (MPI global timer) is defined, the offsets are set to zero. Otherwise, the offsets are measured. Called by MPIB_measure_global. Can be called directly before measurements.

Attention:

As the global variable is an array, the memory should be freed by

 MPIB_global_timer_init(MPI_COMM_NULL, 0) 

Parameters:

	comm	communicator
	parallel	several non-overlapped point-to-point communications at the same time if non-zero
	reps	number of repetitions (not precision because series of ping-pongs are required - we cannot interrupt them by sending/receiving the result of the statistical estimation)

int MPIB_measure_global	(	MPIB_coll_container *	container,
		MPI_Comm	comm,
		int	root,
		int	M,
		MPIB_precision	precision,
		MPIB_result *	result
	)

Measures the execution time of collective operation between processes using global time. Based on . Reuses already obtained offsets between local clocks if the previous initialization was performed on the same communicator. Otherwise, initializes the global timer by calling MPIB_global_timer_init.
In the loop over repetitions:

• Synchronizes the processes by double barrier.
• Measures the moment of start at the root and the moment of finish at the rest of processes.
• Having substructed the offset, finds maximum by reducing to the root.
• Performs statistical analysis at root.

Broadcasts the result.