Classes
class	parsedumpi
	A refactored dumpi parser to read the newer binary-format dumpi trace files. More...

class	parsedumpi_callbacks
	Populate C-style callbacks for a libundumpi parser. More...

class	sumi_transport

Functions
void	comm_init ()

void	comm_finalize ()

int	comm_rank ()

int	comm_nproc ()

void	comm_send_header (int dst, const message::ptr &msg)

void	comm_cancel_ping (int dst, int tag)

void	comm_ping (int dst, int tag, timeout_function *func)

void	comm_send_payload (int dst, const message::ptr &msg)

void	comm_send (int dst, message::payload_type_t ev, const message::ptr &msg)

void	comm_rdma_put (int dst, const message::ptr &msg)

void	comm_rdma_get (int dst, const message::ptr &msg)

void	comm_nvram_get (int dst, const message::ptr &msg)

void	comm_alltoall (void dst, void src, int nelems, int type_size, int tag, bool fault_aware=false, int context=options::initial_context, communicator *dom=0)

void	comm_allgather (void dst, void src, int nelems, int type_size, int tag, bool fault_aware=false, int context=options::initial_context, communicator *dom=0)

void	comm_allgatherv (void dst, void src, int recv_counts, int type_size, int tag, bool fault_aware=false, int context=options::initial_context, communicator dom=0)

void	comm_gather (int root, void dst, void src, int nelems, int type_size, int tag, bool fault_aware=false, int context=options::initial_context, communicator *dom=0)

void	comm_scatter (int root, void dst, void src, int nelems, int type_size, int tag, bool fault_aware=false, int context=options::initial_context, communicator *dom=0)

void	comm_bcast (int root, void buffer, int nelems, int type_size, int tag, bool fault_aware=false, int context=options::initial_context, communicator dom=0)

void	comm_allreduce (void dst, void src, int nelems, int type_size, int tag, reduce_fxn fxn, bool fault_aware=false, int context=options::initial_context, communicator *dom=0)
	The total size of the input/result buffer in bytes is nelems*type_size. More...

template<typename data_t , template< typename > class Op>
void	comm_allreduce (void dst, void src, int nelems, int tag, bool fault_aware=false, int context=options::initial_context, communicator *dom=0)

void	comm_reduce (int root, void dst, void src, int nelems, int type_size, int tag, reduce_fxn fxn, bool fault_aware=false, int context=options::initial_context, communicator *dom=0)

template<typename data_t , template< typename > class Op>
void	comm_reduce (int root, void dst, void src, int nelems, int tag, bool fault_aware=false, int context=options::initial_context, communicator *dom=0)

void	comm_barrier (int tag, bool fault_aware=false, communicator *dom=0)

void	comm_vote (int vote, int tag, vote_fxn fxn, int context=options::initial_context, communicator *dom=0)
	The total size of the input/result buffer in bytes is nelems*type_size This always run in a fault-tolerant fashion This uses a dynamic tree structure that reconnects partners when failures are detected. More...

template<template< class > class VoteOp>
void	comm_vote (int vote, int tag, int context=options::initial_context, communicator *dom=0)

void	comm_kill_node ()
	Helper function. More...

void	comm_kill_process ()
	Helper function. More...

const thread_safe_set< int > &	comm_failed_ranks ()

const thread_safe_set< int > &	comm_failed_ranks (int context)

void	comm_start_heartbeat (double interval)

void	comm_stop_heartbeat ()

collective_done_message::ptr	comm_collective_block (collective::type_t ty, int tag)

message::ptr	comm_poll ()

void	compute (double sec)

void	sleep (double sec)

void	sleep_until (double sec)

int	comm_partner (long node_id)
	Map a physical node location to its virtual assignment in the communicator. More...

double	wall_time ()
	Every node has exactly the same notion of time - universal, global clock. More...

transport *	sumi_api ()

Function Documentation

void sumi::comm_allgather	(	void *	dst,
		void *	src,
		int	nelems,
		int	type_size,
		int	tag,
		bool	fault_aware = `false`,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

void sumi::comm_allgatherv	(	void *	dst,
		void *	src,
		int *	recv_counts,
		int	type_size,
		int	tag,
		bool	fault_aware = `false`,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

void sumi::comm_allreduce	(	void *	dst,
		void *	src,
		int	nelems,
		int	type_size,
		int	tag,
		reduce_fxn	fxn,
		bool	fault_aware = `false`,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

The total size of the input/result buffer in bytes is nelems*type_size.

Parameters

dst	Buffer for the result. Can be NULL to ignore payloads.
src	Buffer for the input. Can be NULL to ignore payloads.
nelems	The number of elements in the input and result buffer.
type_size	The size of the input type, i.e. sizeof(int), sizeof(double)
tag	A unique tag identifier for the collective
fxn	The function that will actually perform the reduction
fault_aware	Whether to execute in a fault-aware fashion to detect failures
context	The context (i.e. initial set of failed procs)

Referenced by comm_allreduce().

template<typename data_t , template< typename > class Op>

void sumi::comm_allreduce	(	void *	dst,
		void *	src,
		int	nelems,
		int	tag,
		bool	fault_aware = `false`,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

Definition at line 103 of file sumi.h.

References comm_allreduce(), and comm_reduce().

Here is the call graph for this function:

void sumi::comm_alltoall	(	void *	dst,
		void *	src,
		int	nelems,
		int	type_size,
		int	tag,
		bool	fault_aware = `false`,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

void sumi::comm_barrier	(	int	tag,
		bool	fault_aware = `false`,
		communicator *	dom = `0`
	)

Referenced by comm_reduce().

void sumi::comm_bcast	(	int	root,
		void *	buffer,
		int	nelems,
		int	type_size,
		int	tag,
		bool	fault_aware = `false`,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

void sumi::comm_cancel_ping	(	int	dst,
		int	tag
	)

collective_done_message::ptr sumi::comm_collective_block	(	collective::type_t	ty,
		int	tag
	)

Referenced by comm_vote().

const thread_safe_set<int>& sumi::comm_failed_ranks ( )

Referenced by comm_vote().

const thread_safe_set<int>& sumi::comm_failed_ranks ( int context )

void sumi::comm_finalize ( )

void sumi::comm_gather	(	int	root,
		void *	dst,
		void *	src,
		int	nelems,
		int	type_size,
		int	tag,
		bool	fault_aware = `false`,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

void sumi::comm_init ( )

void sumi::comm_kill_node ( )

Helper function.

Kill the node that is currently running. This is invoked by an application. This allows an application to die at a very, very specific point in application execution.

Referenced by comm_vote().

void sumi::comm_kill_process ( )

Helper function.

Kill the process that is currently running. This only kills the process - it leaves the node alive and well.

Referenced by comm_vote().

int sumi::comm_nproc ( )

void sumi::comm_nvram_get	(	int	dst,
		const message::ptr &	msg
	)

int sumi::comm_partner ( long node_id )

Map a physical node location to its virtual assignment in the communicator.

Parameters

node_id

Returns

Referenced by comm_vote().

void sumi::comm_ping	(	int	dst,
		int	tag,
		timeout_function *	func
	)

message::ptr sumi::comm_poll ( )

Referenced by comm_vote().

int sumi::comm_rank ( )

void sumi::comm_rdma_get	(	int	dst,
		const message::ptr &	msg
	)

void sumi::comm_rdma_put	(	int	dst,
		const message::ptr &	msg
	)

void sumi::comm_reduce	(	int	root,
		void *	dst,
		void *	src,
		int	nelems,
		int	type_size,
		int	tag,
		reduce_fxn	fxn,
		bool	fault_aware = `false`,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

Referenced by comm_allreduce(), and comm_reduce().

template<typename data_t , template< typename > class Op>

void sumi::comm_reduce	(	int	root,
		void *	dst,
		void *	src,
		int	nelems,
		int	tag,
		bool	fault_aware = `false`,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

Definition at line 115 of file sumi.h.

References comm_barrier(), comm_reduce(), and comm_vote().

Here is the call graph for this function:

void sumi::comm_scatter	(	int	root,
		void *	dst,
		void *	src,
		int	nelems,
		int	type_size,
		int	tag,
		bool	fault_aware = `false`,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

void sumi::comm_send	(	int	dst,
		message::payload_type_t	ev,
		const message::ptr &	msg
	)

void sumi::comm_send_header	(	int	dst,
		const message::ptr &	msg
	)

Parameters

dst	The destination to send to

void sumi::comm_send_payload	(	int	dst,
		const message::ptr &	msg
	)

void sumi::comm_start_heartbeat ( double interval )

Referenced by comm_vote().

void sumi::comm_stop_heartbeat ( )

Referenced by comm_vote().

void sumi::comm_vote	(	int	vote,
		int	tag,
		vote_fxn	fxn,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

The total size of the input/result buffer in bytes is nelems*type_size This always run in a fault-tolerant fashion This uses a dynamic tree structure that reconnects partners when failures are detected.

Parameters

vote	The vote (currently restricted to integer) from this process
nelems	The number of elements in the input and result buffer.
tag	A unique tag identifier for the collective
fxn	The function that merges vote, usually AND, OR, MAX, MIN
context	The context (i.e. initial set of failed procs)

Referenced by comm_reduce(), and comm_vote().

template<template< class > class VoteOp>

void sumi::comm_vote	(	int	vote,
		int	tag,
		int	context = `options::initial_context`,
		communicator *	dom = `0`
	)

Definition at line 138 of file sumi.h.

References comm_collective_block(), comm_failed_ranks(), comm_kill_node(), comm_kill_process(), comm_partner(), comm_poll(), comm_start_heartbeat(), comm_stop_heartbeat(), comm_vote(), compute(), sleep(), sleep_until(), sumi_api(), and wall_time().

Here is the call graph for this function:

void sumi::compute ( double sec )

Referenced by comm_vote(), and sstmac::sw::app::parent_app().

void sumi::sleep ( double sec )

Referenced by comm_vote(), and sstmac::sw::app::parent_app().

void sumi::sleep_until ( double sec )

Referenced by comm_vote().

transport* sumi::sumi_api ( )

Referenced by comm_vote().

double sumi::wall_time ( )

Every node has exactly the same notion of time - universal, global clock.

Thus, if rank 0 starts and 10 minuts later rank 1 starts, even though rank 1 has only been running for 30 seconds, the time will still return 10 mins, 30 seconds.

Returns: The current system wall-clock time in seconds. At application launch, time is zero.

Referenced by comm_vote().

Classes

Functions

Function Documentation