no_mpi.c File Reference

MPI Support Module. More...

#include <distributed/mpi.h>
#include <assert.h>

Include dependency graph for no_mpi.c:

Functions
void	mpi_global_init (int *argc_p, char ***argv_p)
	Initializes the MPI environment. More...
void	mpi_global_fini (void)
	Finalizes the MPI environment.
void	mpi_remote_msg_send (struct lp_msg *msg, nid_t dest_nid)
	Sends a model message to a LP residing on another node. More...
void	mpi_remote_anti_msg_send (struct lp_msg *msg, nid_t dest_nid)
	Sends a model anti-message to a LP residing on another node. More...
void	mpi_control_msg_broadcast (enum msg_ctrl_code ctrl)
	Sends a platform control message to all the nodes, including self. More...
void	mpi_control_msg_send_to (enum msg_ctrl_code ctrl, nid_t dest)
	Sends a platform control message to a specific nodes. More...
void	mpi_remote_msg_handle (void)
	Empties the queue of incoming MPI messages, doing the right thing for each one of them. More...
void	mpi_remote_msg_drain (void)
	Empties the queue of incoming MPI messages, ignoring them. More...
void	mpi_reduce_sum_scatter (const uint32_t values[n_nodes], uint32_t *result)
	Computes the sum-reduction-scatter operation across all nodes. More...
bool	mpi_reduce_sum_scatter_done (void)
	Checks if a previous mpi_reduce_sum_scatter() operation has completed. More...
void	mpi_reduce_min (double *node_min_p)
	Computes the min-reduction operation across all nodes. More...
bool	mpi_reduce_min_done (void)
	Checks if a previous mpi_reduce_min() operation has completed. More...
void	mpi_node_barrier (void)
	A node barrier.
void	mpi_blocking_data_send (const void *data, int data_size, nid_t dest)
	Sends a byte buffer to another node. More...
void *	mpi_blocking_data_rcv (int *data_size_p, nid_t src)
	Receives a byte buffer from another node. More...

Detailed Description

MPI Support Module.

This module implements all basic MPI facilities to let the distributed execution of a simulation model take place consistently.

Several facilities are thread-safe, others are not. Check carefully which of these can be used by worker threads without coordination when relying on this module.

Copyright

Copyright (C) 2008-2021 HPDCS Group https://hpdcs.github.io

Function Documentation

mpi_blocking_data_rcv()

void* mpi_blocking_data_rcv	(	int *	data_size_p,
		nid_t	src
	)

Receives a byte buffer from another node.

Parameters

data_size_p	where to write the size of the received data
src	the id of the sender node

Returns

the buffer allocated with mm_alloc() containing the received data

This operation blocks the execution until the sender node actually sends the data with mpi_raw_data_blocking_send().

mpi_blocking_data_send()

void mpi_blocking_data_send	(	const void *	data,
		int	data_size,
		nid_t	dest
	)

Sends a byte buffer to another node.

Parameters

data	a pointer to the buffer to send
data_size	the buffer size
dest	the id of the destination node

This operation blocks the execution flow until the destination node receives the data with mpi_raw_data_blocking_rcv().

mpi_control_msg_broadcast()

void mpi_control_msg_broadcast

(

enum msg_ctrl_code

ctrl

)

Sends a platform control message to all the nodes, including self.

Parameters

ctrl	the control message to send

mpi_control_msg_send_to()

void mpi_control_msg_send_to	(	enum msg_ctrl_code	ctrl,
		nid_t	dest
	)

Sends a platform control message to a specific nodes.

Parameters

ctrl	the control message to send
dest	the id of the destination node

mpi_global_init()

void mpi_global_init	(	int *	argc_p,
		char ***	argv_p
	)

Initializes the MPI environment.

Parameters

argc_p	a pointer to the OS supplied argc
argv_p	a pointer to the OS supplied argv

mpi_reduce_min()

void mpi_reduce_min

(

double *

node_min_p

)

Computes the min-reduction operation across all nodes.

Parameters

node_min_p

a pointer to the value from the calling node which will also be used to store the computed minimum.

Each node supplies a single simtime_t value. The minimum of all these values is computed and stored in node_min_p itself. It is expected that only a single thread calls this function at a time. Each node has to call this function else the result can't be computed. It is possible to have a single mpi_reduce_min() operation pending at a time. Both arguments must point to valid memory regions until mpi_reduce_min_done() returns true.

mpi_reduce_min_done()

bool mpi_reduce_min_done

(

void

)

Checks if a previous mpi_reduce_min() operation has completed.

Returns

true if the previous operation has been completed, false otherwise.

mpi_reduce_sum_scatter()

void mpi_reduce_sum_scatter	(	const uint32_t	values[n_nodes],
		uint32_t *	result
	)

Computes the sum-reduction-scatter operation across all nodes.

Parameters

values	a flexible array implementing the addendum vector from the calling node.
result	a pointer where the nid-th component of the sum will be stored.

Each node supplies a n_nodes components vector. The sum of all these vector is computed and the nid-th component of this vector is stored in result. It is expected that only a single thread calls this function at a time. Each node has to call this function else the result can't be computed. It is possible to have a single mpi_reduce_sum_scatter() operation pending at a time. Both arguments must point to valid memory regions until mpi_reduce_sum_scatter_done() returns true.

mpi_reduce_sum_scatter_done()

bool mpi_reduce_sum_scatter_done

(

void

)

Checks if a previous mpi_reduce_sum_scatter() operation has completed.

Returns

true if the previous operation has been completed, false otherwise.

mpi_remote_anti_msg_send()

void mpi_remote_anti_msg_send	(	struct lp_msg *	msg,
		nid_t	dest_nid
	)

Sends a model anti-message to a LP residing on another node.

Parameters

msg	the message to rollback
dest_nid	the id of the node where the targeted LP resides

This function also calls the relevant handlers in order to keep, for example, the non blocking gvt algorithm running. Note that when this function returns, the anti-message may have not been sent yet. We don't need to actively check for sending completion: the platform, during the fossil collection, leverages the gvt to make sure the message has been indeed sent and processed before freeing it.

mpi_remote_msg_drain()

void mpi_remote_msg_drain

(

void

)

Empties the queue of incoming MPI messages, ignoring them.

This routine checks, using the MPI probing mechanism, for new remote messages and it discards them. It is used at simulation completion to clear MPI state.

mpi_remote_msg_handle()

void mpi_remote_msg_handle

(

void

)

Empties the queue of incoming MPI messages, doing the right thing for each one of them.

This routine checks, using the MPI probing mechanism, for new remote messages and it handles them accordingly. Control messages are handled by the respective platform handler. Simulation messages are unpacked and put in the queue. Anti-messages are matched and accordingly processed by the message map.

mpi_remote_msg_send()

void mpi_remote_msg_send	(	struct lp_msg *	msg,
		nid_t	dest_nid
	)

Sends a model message to a LP residing on another node.

Parameters

msg	the message to send
dest_nid	the id of the node where the targeted LP resides

This function also calls the relevant handlers in order to keep, for example, the non blocking gvt algorithm running. Note that when this function returns, the message may have not been actually sent. We don't need to actively check for sending completion: the platform, during the fossil collection, leverages the gvt to make sure the message has been indeed sent and processed before freeing it.