Inheritance diagram for PV::Checkpointer:

Data Structures
struct	TimeInfo

Checkpointer Parameters
List of parameters needed from the Checkpointer class
virtual void	ioParam_verifyWrites (enum ParamsIOFlag ioFlag, PVParams *params)
	verifyWrites: If true, calls to FileStream::write are checked by opening the file in read mode and reading back the data and comparing it to the data just written.

virtual void	ioParam_outputPath (enum ParamsIOFlag ioFlag, PVParams *params)
	mOutputPath: Specifies the absolute or relative output path of the run

void	ioParam_checkpointWrite (enum ParamsIOFlag ioFlag, PVParams *params)
	checkpointWrite: Flag to determine if the run writes checkpoints.

void	ioParam_checkpointWriteDir (enum ParamsIOFlag ioFlag, PVParams *params)
	checkpointWriteDir: If checkpointWrite is set, specifies the output checkpoint directory.

void	ioParam_checkpointWriteTriggerMode (enum ParamsIOFlag ioFlag, PVParams *params)
	mCheckpointWriteTriggerMode: If checkpointWrite is set, specifies the method to checkpoint. More...

void	ioParam_checkpointWriteStepInterval (enum ParamsIOFlag ioFlag, PVParams *params)
	checkpointWriteStepInterval: If checkpointWrite on step, specifies the number of steps between checkpoints.

void	ioParam_checkpointWriteTimeInterval (enum ParamsIOFlag ioFlag, PVParams *params)
	checkpointWriteTimeInteval: If checkpointWrite on time, specifies the amount of simulation time between checkpoints.

void	ioParam_checkpointWriteClockInterval (enum ParamsIOFlag ioFlag, PVParams *params)
	checkpointWriteClockInteval: If checkpointWrite on clock, specifies the amount of clock time between checkpoints. The units are specified using the parameter checkpointWriteClockUnit

void	ioParam_checkpointWriteClockUnit (enum ParamsIOFlag ioFlag, PVParams *params)
	checkpointWriteClockInteval: If checkpointWrite on clock, specifies the units used in checkpointWriteClockInterval.

void	ioParam_checkpointIndexWidth (enum ParamsIOFlag ioFlag, PVParams *params)
	If checkpointWrite is true, checkpointIndexWidth specifies the minimum width for the step number appearing in the checkpoint directory. More...

void	ioParam_suppressNonplasticCheckpoints (enum ParamsIOFlag ioFlag, PVParams *params)

void	ioParam_deleteOlderCheckpoints (enum ParamsIOFlag ioFlag, PVParams *params)
	deleteOlderCheckpoints: If checkpointWrite, specifies if the run should delete older checkpoints when writing new ones.

void	ioParam_numCheckpointsKept (enum ParamsIOFlag ioFlag, PVParams *params)
	mNumCheckpointsKept: If mDeleteOlderCheckpoints is set, keep this many checkpoints before deleting the checkpoint. Default is 1 (delete a checkpoint when a newer checkpoint is written.)

void	ioParam_initializeFromCheckpointDir (enum ParamsIOFlag ioFlag, PVParams *params)
	initializeFromCheckpointDir: Sets directory used by Checkpointer::initializeFromCheckpoint(). Layers and connections use this directory if they set their initializeFromCheckpointFlag parameter.

void	ioParam_lastCheckpointDir (enum ParamsIOFlag ioFlag, PVParams *params)
	lastCheckpointDir: If checkpointWrite is not set, this required parameter specifies the directory to write a final written checkpoint at the end of the run. Writing the last checkpoint can be suppressed by setting this string to the empty string. Relative paths are relative to the working directory.

Public Member Functions
	Checkpointer (std::string const &name, MPIBlock const globalMPIBlock, Arguments const arguments)

virtual void	addObserver (Observer *observer) override

void	checkpointRead (double simTimePointer, long int currentStepPointer)

void	checkpointWrite (double simTime)

bool	doesVerifyWrites ()

void	finalCheckpoint (double simTime)

std::string const &	getBlockDirectoryName () const

std::string const &	getCheckpointReadDirectory () const

char const *	getCheckpointWriteDir () const

bool	getCheckpointWriteFlag () const

double	getCheckpointWriteSimtimeInterval () const

long int	getCheckpointWriteStepInterval () const

enum CheckpointWriteTriggerMode	getCheckpointWriteTriggerMode () const

char const *	getInitializeFromCheckpointDir () const

char const *	getLastCheckpointDir () const

MPIBlock const *	getMPIBlock ()

std::string const &	getOutputPath ()

bool	getSuppressNonplasticCheckpoints () const

void	ioParams (enum ParamsIOFlag ioFlag, PVParams *params)

std::string	makeOutputPathFilename (std::string const &path)

void	provideFinalStep (long int finalStep)

void	readNamedCheckpointEntry (std::string const &objName, std::string const &dataName, bool constantEntireRun)

void	readNamedCheckpointEntry (std::string const &checkpointEntryName, bool constantEntireRun=false)

void	readStateFromCheckpoint ()

template<typename T >
bool	registerCheckpointData (std::string const &objName, std::string const &dataName, T *dataPointer, size_t numValues, bool broadcast, bool constantEntireRun)

bool	registerCheckpointEntry (std::shared_ptr< CheckpointEntry > checkpointEntry, bool constantEntireRun)

void	registerTimer (Timer const *timer)

void	writeTimers (PrintStream &stream) const

Protected Member Functions
Response::Status	notify (ObserverTable const &table, std::vector< std::shared_ptr< BaseMessage const >> messages, bool printFlag)

Response::Status	notify (ObserverTable const &table, std::shared_ptr< BaseMessage const > message, bool printFlag)

void	notifyLoop (ObserverTable const &table, std::vector< std::shared_ptr< BaseMessage const >> messages, bool printFlag, std::string const &description)

void	notifyLoop (ObserverTable const &table, std::shared_ptr< BaseMessage const > message, bool printFlag, std::string const &description)

Private Types
enum	CheckpointWriteTriggerMode { NONE, STEP, SIMTIME, WALLCLOCK }

enum	WallClockUnit { SECOND, MINUTE, HOUR, DAY }

Private Member Functions
void	checkpointNow ()

void	checkpointToDirectory (std::string const &checkpointDirectory)

void	checkpointWriteSignal ()

void	extractCheckpointReadDirectory ()

void	findWarmStartDirectory ()

std::string	generateBlockPath (std::string const &baseDirectory)

void	initBlockDirectoryName ()

void	initMPIBlock (MPIBlock const globalMPIBlock, Arguments const arguments)

void	ioParamsFillGroup (enum ParamsIOFlag ioFlag, PVParams *params)

std::string	makeCheckpointDirectoryFromCurrentStep ()

bool	receivedSignal ()

void	rotateOldCheckpoints (std::string const &newCheckpointDirectory)

bool	scheduledCheckpoint ()

bool	scheduledSimTime ()

bool	scheduledStep ()

bool	scheduledWallclock ()

void	verifyDirectory (char const *directory, std::string const &description)

void	writeTimers (std::string const &directory)

Private Attributes
std::string	mBlockDirectoryName

int	mCheckpointIndexWidth = -1

std::string	mCheckpointReadDirectory

std::vector< std::shared_ptr< CheckpointEntry > >	mCheckpointRegistry

Timer *	mCheckpointTimer = nullptr

char *	mCheckpointWriteDir = nullptr

bool	mCheckpointWriteFlag = false

double	mCheckpointWriteSimtimeInterval = 1.0

long int	mCheckpointWriteStepInterval = 1L

enum CheckpointWriteTriggerMode	mCheckpointWriteTriggerMode = NONE

char *	mCheckpointWriteTriggerModeString = nullptr

std::time_t	mCheckpointWriteWallclockInterval = 1L

std::time_t	mCheckpointWriteWallclockIntervalSeconds = 1L

char *	mCheckpointWriteWallclockUnit = nullptr

bool	mDeleteOlderCheckpoints = false

char *	mInitializeFromCheckpointDir = nullptr

char *	mLastCheckpointDir = nullptr

std::time_t	mLastCheckpointWallclock = (std::time_t)0

MPIBlock *	mMPIBlock = nullptr

std::string	mName

double	mNextCheckpointSimtime = 0.0

long int	mNextCheckpointStep = 0L

std::time_t	mNextCheckpointWallclock = (std::time_t)0

int	mNumCheckpointsKept = 2

ObserverTable	mObserverTable

std::vector< std::string >	mOldCheckpointDirectories

int	mOldCheckpointDirectoriesIndex

std::string	mOutputPath = ""

bool	mSuppressNonplasticCheckpoints = false

TimeInfo	mTimeInfo

std::shared_ptr< CheckpointEntryData< TimeInfo > >	mTimeInfoCheckpointEntry = nullptr

std::vector< Timer const * >	mTimers

bool	mVerifyWrites = true

bool	mWarmStart = false

int	mWidthOfFinalStepNumber = 0

Static Private Attributes
static std::string const	mDefaultOutputPath = "output"

Detailed Description

Definition at line 25 of file Checkpointer.hpp.

Member Function Documentation

void PV::Checkpointer::addObserver ( Observer * observer )

overridevirtual

The virtual method for adding an Observer-derived object. Derived classes must override this method to add the object to their hierarchy.

Reimplemented from PV::Subject.

Definition at line 456 of file Checkpointer.cpp.

void PV::Checkpointer::checkpointNow ( )

private

Called by checkpointWrite() (if there was not a SIGUSR1 signal pending) and finalCheckpoint(). Writes a checkpoint, indexed by the current timestep. If the deleteOlderCheckpoints param was set, and the number of checkpoints exceeds numCheckpointsKept, the oldest checkpoint is deleted, and the just-written checkpoint is rotated onto the list of checkpoints that will be deleted.

Definition at line 772 of file Checkpointer.cpp.

void PV::Checkpointer::checkpointToDirectory ( std::string const & checkpointDirectory )

private

Creates a checkpoint based at the given directory. If the checkpoint directory already exists, it issues a warning, and deletes the timeinfo.bin file in the checkpooint. This way, the presence of the timeinfo.bin file indicates that the checkpoint is complete.

Definition at line 798 of file Checkpointer.cpp.

void PV::Checkpointer::checkpointWriteSignal ( )

private

Called by checkpointWrite if a SIGUSR1 signal was sent (as reported by receivedSignal). It writes a checkpoint, indexed by the current timestep. If the deleteOlderCheckpoints param was set, it does not cause a checkpoint to be deleted, and does not rotate the checkpoint into the list of directories that will be deleted.

Definition at line 752 of file Checkpointer.cpp.

void PV::Checkpointer::extractCheckpointReadDirectory ( )

private

If called when mCheckpointReadDirectory is a colon-separated list of paths, extracts the entry corresponding to the process's batch index and replaces mCheckpointReadDirectory with that entry. Called by configCheckpointReadDirectory. If mCheckpointReadDirectory is not a colon-separated list, it is left unchanged. If it is a colon-separated list, the number of entries must agree with

Definition at line 595 of file Checkpointer.cpp.

void PV::Checkpointer::ioParam_checkpointIndexWidth	(	enum ParamsIOFlag	ioFlag,
		PVParams *	params
	)

private

If checkpointWrite is true, checkpointIndexWidth specifies the minimum width for the step number appearing in the checkpoint directory.

If the step number needs fewer digits than checkpointIndexWidth, it is padded with zeroes. If the step number needs more, the full step number is still printed. Hence, setting checkpointWrite to zero means that there are never any padded zeroes. If set to a negative number, the width will be inferred from startTime, stopTime and dt. The default value is -1 (infer the width).

Definition at line 402 of file Checkpointer.cpp.

void PV::Checkpointer::ioParam_checkpointWriteTriggerMode	(	enum ParamsIOFlag	ioFlag,
		PVParams *	params
	)

private

mCheckpointWriteTriggerMode: If checkpointWrite is set, specifies the method to checkpoint.

Possible choices include

step: Checkpoint off of timesteps
time: Checkpoint off of simulation time
clock: Checkpoint off of clock time. Not implemented yet.

Definition at line 179 of file Checkpointer.cpp.

void PV::Checkpointer::ioParam_suppressNonplasticCheckpoints	(	enum ParamsIOFlag	ioFlag,
		PVParams *	params
	)

private

If checkpointWrite is true and this flag is true, connections will only checkpoint if plasticityFlag=true.

Definition at line 414 of file Checkpointer.cpp.

std::string PV::Checkpointer::makeOutputPathFilename ( std::string const & path )

Given a relative path, returns a full path consisting of the effective output directory for the process's checkpoint cell, followed by "/", followed by the given relative path. It is a fatal error for the path to be an absolute path (i.e. starting with '/').

Definition at line 87 of file Checkpointer.cpp.

Response::Status PV::Subject::notify	(	ObserverTable const &	table,
		std::vector< std::shared_ptr< BaseMessage const >>	messages,
		bool	printFlag
	)

protectedinherited

This method calls the respond() method of each object in the given table, using the given vector of messages. If the table consists of objects A, B, and C; and the messages vector consists of messages X and Y, the order is A->X, A->Y, B->X, B->Y, C->X, C->Y.

The objects' respond() method returns one of the Response::Status types: SUCCESS, NO_ACTION, PARTIAL, or PV_POSTPONE.

SUCCESS: the object completed the task requested by the messages. NO_ACTION: the object has nothing to do in response to the message, or had already done it. PARTIAL: the object has not yet completed but is making progress. It is expected that there are a small number of descrete tasks, so that an object will not return PARTIAL a large number of times in response to the same message. POSTPONE: the object needs to act but cannot do so until an event outside its control occurs.

If all objects return NO_ACTION, then notify() returns NO_ACTION. If all objects return either SUCCESS or NO_ACTION and there is at least one SUCCESS, then notify() returns SUCCESS. If all objects return either POSTPONE or NO_ACTION and there is at least one POSTPONE, then notify() returns POSTPONE. Otherwise, notify() returns PARTIAL.

Generally each message in the messages vector is sent to each object in the table. However, if an object returns POSTPONE in response to a message, the loop skips to the next object, and does not sent any remaining messages to the postponing object.

The rationale behind these rules is so that if the objects in the table are themselves derived from the Subject class, the messages can be passed down the tree and the return values passed up it, and the return value at the top can be interpreted as being over all the components at the bottom, without the topmost object needing to know details of the composition of the objects below it.

If printFlag is true, the method prints information regarding postponement to standard output.

Definition at line 15 of file Subject.cpp.

Response::Status PV::Subject::notify	(	ObserverTable const &	table,
		std::shared_ptr< BaseMessage const >	message,
		bool	printFlag
	)

inlineprotectedinherited

A convenience overload of the basic notify method where there is only one message to send to the objects. This overloading handles enclosing the message in a vector of length one.

Definition at line 91 of file Subject.hpp.

void PV::Subject::notifyLoop	(	ObserverTable const &	table,
		std::vector< std::shared_ptr< BaseMessage const >>	messages,
		bool	printFlag,
		std::string const &	description
	)

protectedinherited

This method calls the notify() method in a loop until the result is not PARTIAL. If it is either POSTPONE, it exits with a fatal error. The description argument is used in the error message to report which message vector failed. If the result is PV_SUCCESS, notifyLoop() returns to the calling function.

notifyLoop should only be used if the objects that might cause a postponement are themselves in the table of objects; otherwise the routine will hang.

Definition at line 57 of file Subject.cpp.

void PV::Subject::notifyLoop	(	ObserverTable const &	table,
		std::shared_ptr< BaseMessage const >	message,
		bool	printFlag,
		std::string const &	description
	)

inlineprotectedinherited

A convenience overload of the basic notifyLoop method where there is only one message to send to the objects. This overloading handles enclosing the message in a vector of length one.

Definition at line 114 of file Subject.hpp.

bool PV::Checkpointer::receivedSignal ( )

private

If a SIGUSR signal has been received by the global root process, clears the signal and returns true. Otherwise returns false.

Definition at line 676 of file Checkpointer.cpp.

void PV::Checkpointer::rotateOldCheckpoints ( std::string const & newCheckpointDirectory )

private

Called if deleteOlderCheckpoints is true. It deletes the oldest checkpoint in the list of old checkpoint directories, and adds the new checkpoint directory to the list.

Definition at line 843 of file Checkpointer.cpp.

bool PV::Checkpointer::scheduledCheckpoint ( )

private

Returns true if the params file settings indicate a checkpoint should occur at this timestep. It also advances the appropriate data member for the trigger mode to the next scheduled checkpoint. Returns false otherwise.

Definition at line 700 of file Checkpointer.cpp.

bool PV::Checkpointer::scheduledSimTime ( )

private

Called by scheduledCheckpoint if checkpointWriteTriggerMode is "time". If the simTime is >= the current value of mNextCheckpointSimtime, it advances mNextCheckpointSimtime by the time interval and returns true. Otherwise it returns false.

Definition at line 725 of file Checkpointer.cpp.

bool PV::Checkpointer::scheduledStep ( )

private

Called by scheduledCheckpoint if checkpointWriteTriggerMode is "step". If the step number is an integral multiple of checkpointWriteStepInterval, it advances mNextCheckpointStep by the step interval and returns true. Otherwise it returns false.

Definition at line 715 of file Checkpointer.cpp.

bool PV::Checkpointer::scheduledWallclock ( )

private

Called by scheduledCheckpoint if checkpointWriteTriggerMode is "clock". If the elapsed time between the wall clock time and mLastCheckpointWallclock exceeds mCheckpointWriteWallclockInterval, it sets mLastCheckpointWallclock to the current wall clock time and returns true. Otherwise it returns false.

Definition at line 734 of file Checkpointer.cpp.

Field Documentation

int PV::Checkpointer::mOldCheckpointDirectoriesIndex

private

Initial value:

=

0

Definition at line 334 of file Checkpointer.hpp.

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

OpenPV/src/checkpointing/Checkpointer.hpp
OpenPV/src/checkpointing/Checkpointer.cpp

Data Structures

Checkpointer Parameters

Public Member Functions

Protected Member Functions

Private Types

Private Member Functions

Private Attributes

Static Private Attributes

Detailed Description

Member Function Documentation

Field Documentation