From GridLAB-D Wiki
Jump to: navigation, search

Creating a module - Procedure to create a new module in GridLAB-D

Runtime modules are implemented as dynamic libraries that are loaded as needed. The model loaded determines whether a runtime module is needed by specifying a module block.

Prior to Diablo (Version 2.0)
Adding a module in Windows can be done using the "Add GridLAB module" wizard. In Linux it can be done using the "add_gridlab_module" script. This document is provided for completeness and to provide details that may be necessary should the scripts not function as required.

The exact approach to use in building a GridLAB module is not clear cut. In general a module is a solver that can compute the steady state of a collection of objects given a specific boundary condition. For example, a power flow solver makes sense as a module because the steady state of a flow network can be directly computed. However, a market clearing system and a load simulation doesn't make sense because the market is influenced not only by demand from loads, but also by supply. As a general rule, if a set of simultaneous equations can be solved to obtain the state of a system, the system is suitable for implementation as a module.

Modules must be able to implement a least three capabilities:

  • they must be able to create objects on demand (see create)
  • they must be able to initialize objects on demand (see init)
  • they must be able to compute the state of individual objects at a specified date and time on demand (see sync)

In addition, modules generally should be able to implement the following

  • compute the states of objects with some degree of parallelism
  • load and save data in various exchange formats
  • inform the GridLAB-D core which objects data members are exposed to other modules
  • handle notification events of data members about to be changed or just changed
  • determine whether an object is a subtype of another object
  • verify that a collection of objects form a self-consistent and correct model

It turns out that implementing these capabilities is not as easy as it at first seems. In particular, the synchronization has typically been one of the most challenging concepts for programmers to understand. Given the amount of time spend in sync calls, it is recommended that considerable time and effort be put into its design.

Basic Synchronization

An object's sync method actually performs two essential functions. First, it updates the state of an object to a designated point in time, and second it lets the core know when the object is next expected to change state. This is vital for the core to know because the core's clock will be advanced to the time of the next expected state change, and all objects will be synchronized to that time.

In general a sync() function should be aware of three time values:

  • <math>t_0</math> is the time from which the object is being advanced. This is not the current time, because it is presumed that the object has not yet advanced to the current time and this is why sync() is being called.
  • <math>t_1</math> is the time to which the object is being advanced. Think

of this as \e now from the object's perspective. This is usually the current time from the core's perspective (but don't assume it \e always is).

  • <math>t_2</math> is the time at which the object expects to have its next

state change. This time must be computed by the object during or immediately after the state is advanced to <math>t_1</math>. This is the time returned to the core should the sync() call succeed.

If no state change is ever anticipated then <math>t_2 = TS\_NEVER</math> is returned, indicating that barring any changes to its boundary condition, the object is in steady state.

If an object's sync() method determines that the object is not yet in steady state (i.e., the module has not converged), then <math>t_2 = t_1</math> is returned.

If an object's sync() method determines that it cannot update to <math>t_1</math> as required, the simulation has failed. It can either throw an exception using GL_THROW() or return <math>t_2 < t_1</math> to indicate the time at which the problem is believed to have occurred.

The time window <math>t_1 - t_0</math> is the past window and the sync() method must implement all behaviors during that time as though they have already occurred.

The time window <math>t_2 - t_1</math> is the future window and the sync function must not implement behaviors in this window yet, as they have not yet occurred.

It is a non-trivial fact that if all objects in all modules in GridLAB model return <math>t_2 = TS\_NEVER</math>, then the simulation stops successfully. This is because the system has completely settle to a steady state and there is nothing further to compute.

For more details on synchronization in GridLAB-D see theory of operation.

Control Code

One very important aspect of synchronization behavior is how control code is handled when object behavior goes beyond the mere physics of its response to its boundary condition. It is quite easy to implement control code that is integrated with the physical model. However, this would prevent users from altering the control code without altering the source code of the object's implementation.

Prior to Hassayampa (Version 3.0)
To address this problem, objects can implement default control code that is disabled if a PLC object is attached later. The ability to alter control code should be made available when implemented for any object for which this is a realistic possibility, which is very nearly always.
To implement default machine PLC code for an object, the module must expose a plc() method that will be called immediately before sync() is called, but only if not external PLC method is provided in the model. This plc() method may be written as though it was integrated with the physics implemented in sync(), but the physics must be able to update even when the default PLC code is not run.
Hassayampa (Version 3.0) and later
As of Hassayampa (Version 3.0) the PLC module will no longer be supported and the PID controller module replaces it. Alternatively, the class intrinsic function plc supports simple control code replacements.

Building a GridLAB module

A GridLAB module must be a Windows DLL or a Linux SO.

VS2005

The compile options that are typically required are

  • Include path: ../core
  • Runtime library: Multi-threaded Debug DLL (/MDd)

For example the command-line options used by the climate module is

compiler: /Od /I "../core" /I "../test/cppunit2/include"
	/D "WIN32" /D "_DEBUG" /D "_WINDOWS" /D "_USRDLL"
		/D "_CRT_SECURE_NO_DEPRECATE" /D "_WINDLL" /D "_MBCS"
		/Gm /EHsc /RTC1 /MDd /Fo"Win32\Debug\climate\"
		/Fd"Win32\Debug\climate\vc80.pdb" /W3 /nologo /c /Wp64
		/ZI /TP /errorReport:prompt

	linker: /OUT:"Win32\Debug\climate.dll" /INCREMENTAL /NOLOGO
	/LIBPATH:"..\test\cppunit2\lib" /DLL /MANIFEST
	/MANIFESTFILE:"Win32\Debug\climate\climate.dll.intermediate.manifest"
	/DEBUG /PDB:"c:\projects\GridlabD\source\VS2005\Win32\Debug\climate.pdb"
	/SUBSYSTEM:WINDOWS /MACHINE:X86 /ERRORREPORT:PROMPT cppunitd.lib
	kernel32.lib user32.lib gdi32.lib winspool.lib comdlg32.lib
	advapi32.lib shell32.lib ole32.lib oleaut32.lib uuid.lib odbc32.lib odbccp32.lib

Linux makefiles

configure.ac
You must add the module's makefile to the AC_CONFIG_FILES variable near the end of the list (but not in the last place).
Makefile.am
You must add the module folder location to the SUBDIRS variable in the last position.

Module implementation Prior to Hassayampa (Version 3.0)

The main.cpp file contains the code to load and unload both Windows DLL and Linux/Unix shared-object libraries:

	// the version can be used to make sure the right module is loaded
	#define MAJOR 0 // TODO: set the major version of your module here
	#define MINOR 0 // TODO: set the minor version of your module here

	#define DLMAIN
	#include <stdlib.h>
	#include "gridlabd.h"
	EXPORT int do_kill(void*);
	EXPORT int major=MAJOR, minor=MINOR;

	#ifdef WIN32
	#define WIN32_LEAN_AND_MEAN		// Exclude rarely-used stuff from Windows headers
	#include <windows.h>
	BOOL APIENTRY DllMain( HANDLE hModule, DWORD  ul_reason_for_call, LPVOID)
	{
		switch (ul_reason_for_call)
		{
			case DLL_PROCESS_ATTACH:
			case DLL_THREAD_ATTACH:
				break;
			case DLL_THREAD_DETACH:
			case DLL_PROCESS_DETACH:
				do_kill(hModule);
				break;
		}
		return TRUE;
	}
	#else // !WIN32
	CDECL int dllinit() __attribute__((constructor));
	CDECL int dllkill() __attribute__((destructor));
	CDECL int dllinit() { return 0;}
	CDECL int dllkill() { do_kill(NULL);}
	#endif // !WIN32
	\endcode

The init.cpp file contains the code needed to initialize a module after it is loaded:

	#include <stdlib.h>
	#include <stdio.h>
	#include <errno.h>
	#include "gridlabd.h"
	#include "myclass.h"
	EXPORT CLASS *init(CALLBACKS *fntable, MODULE *module, int argc, char *argv[])
	{
		// set the GridLAB core API callback table
		callback = fntable;

		// TODO: register each object class by creating its first instance
		new myclass(module);

		// always return the first class registered
		return myclass::oclass;
	}
	CDECL int do_kill()
	{
		// if anything needs to be deleted or freed, this is a good time to do it
		return 0;
	}

You can also implement EXPORT int check(void) to allow the core to request a check of the objects that are implemented by the module. This is particularly useful to perform topology checks for network models.

If you implement the EXPORT int import_file(char *file) this will permit users to use the import command in the model definition.

If you implement the EXPORT int save(char *file) this will permit users to request saves in formats other than .glm or .xml.

Prior to Diablo (Version 2.0)
The EXPORT int setvar(char *varname, char *value) and the

EXPORT void* getvar(char *varname, char *value, unsigned int size) are implemented when you wish to allow the user to alter any of the module's global variables. See network/varmap.c for an example.

EXPORT int module_test(TEST_CALLBACKS *callbacks,int argc, char* argv[])

function implements the module's unit test code. See "Unit testing" for more information.

The EXPORT MODULE *subload(char *modname, MODULE **pMod, CLASS **cptr, int argc, char **argv) function is used to load modules written in foreign languages. Look at the gldjava project for examples of how this is used. This function needs to manually set the function pointers for any classes registered by subload-ed modules. A module subload is triggered with "module X::Y".

The implementation of each class will require two files for each object class be included in your module's source code. The header file will usually include the following:

	#ifndef _MYCLASS_H
	#define _MYCLASS_H
	#include "gridlabd.h"
	class myclass {
	public:
		// TODO: add your published variables here using GridLAB types (see PROPERTY)
		double myDouble; // just an example
	private:
		// TODO: add any unpublished variables here (any type is ok)
		double *pDouble; // just an example
	public:
		static CLASS *oclass;
		static myclass *defaults;
	public:
		myclass(MODULE *module);
		int create(void);
		int init(OBJECT *parent);
		TIMESTAMP sync(TIMESTAMP t0, TIMESTAMP t1);
	};
	#endif

The implementation file should include the following:

	#include <stdlib.h>
	#include <stdio.h>
	#include <errno.h>
	#include "myclass.h"
	CLASS *myclass::oclass = NULL;
	myclass *myclass::defaults = NULL;
	myclass::myclass(MODULE *module)
	{
		if (oclass==NULL)
		{
			oclass = gl_register_class(module,"myclass",PC_BOTTOMUP);
			if (gl_publish_variable(oclass,
				// TODO: publish your variables here
				PT_double, "my_double", PADDR(myDouble), // just an example
				NULL)<1) GL_THROW("unable to publish properties in %s",__FILE__);
			defaults = this;
			// TODO: initialize the default values that apply to all objects of this class
			myDouble = 1.23; // just an example
			pDouble = NULL; // just an example
		}
	}
	void myclass::create(void)
	{
		memcpy(this,defaults,sizeof(*this));
		// TODO: initialize the defaults values that do not depend on relatioships with other objects
	}
	int myclass::init(OBJECT *parent)
	{
		// find the data in the parent object
		myclass *pMyClass = OBJECTDATA(parent,myclass); // just an example

		// TODO: initialize the default values that depend on relationships with other objects
		pDouble = &(pMyClass->double); // just an example

		// return 1 on success, 0 on failure
		return 1;
	}
	TIMESTAMP myclass::sync(TIMESTAMP t0, TIMESTAMP t1)
	{
		// TODO: update the state of the object
		myDouble = myDouble*1.01; // just an example

		// TODO: compute time to next state change
		return (TIMESTAMP)(t1 + myDouble/TS_SECOND); // just an example
	}

	//////////////////////////////////////////////////////////////////////////
	// IMPLEMENTATION OF CORE LINKAGE
	//////////////////////////////////////////////////////////////////////////

	/// Create a myclass object
	EXPORT int create_myclass(OBJECT **obj, ///< a pointer to the OBJECT*
							  OBJECT *parent) ///< a pointer to the parent OBJECT
	{
		*obj = gl_create_object(myclass::oclass,sizeof(myclass));
		if (*obj!=NULL)
		{
			myclass *my = OBJECTDATA(*obj,myclass);
			gl_set_parent(*obj,parent);
			my->create();
			return 1;
		}
		return 0;
	}

	/// Initialize the object
	EXPORT int init_myclass(OBJECT *obj, ///< a pointer to the OBJECT
							OBJECT *parent) ///< a pointer to the parent OBJECT
	{
		if (obj!=NULL)
		{
			myclass *my = OBJECTDATA(obj,myclass);
			return my->init(parent);
		}
		return 0;
	}

	/// Synchronize the object
	EXPORT TIMESTAMP sync_myclass(OBJECT *obj, ///< a pointer to the OBJECT
				  TIMESTAMP t1) ///< the time to which the OBJECT's clock should advance
	{
		TIMESTAMP t2 = OBJECTDATA(obj,myclass)->sync(obj->clock,t1);
		obj->clock = t1;
		return t2;
	}

There are a number of useful extended capabilities that can be added. These include

  • int notify_myclass(OBJECT *obj, NOTIFYMODULE msg)

can be implemented to receive notification messages before and after a variable is changed by the core (NM_PREUPDATE / NM_POSTUPDATE) and when the core needs the module to reset (NM_RESET)

  • int isa_myclass(OBJECT *obj, char *classname)

can be implemented to allow the core to discover whether an object is a subtype of another class.

  • int plc_myclass(OBJECT *obj, TIMESTAMP t1)

can be implemented to create default PLC code that can be overridden by attaching a child plc object.

  • EXPORT int recalc_myclass(OBJECT *obj)

can be implemented to create a recalculation triggered based on changes to properties made through object_set_value_by_addr() and related functions. A property can be made to trigger recalculation calls by adding PT_FLAGS, PF_RECALC after the property publish specification, e.g.,

gl_publish_variable(oclass, PT_double, "resistance[Ohm]", PADDR(resistance), PT_FLAGS, PF_RECALC, NULL);

will cause recalc() to be called after resistance is changed. The recalc calls occur right before the PLC code is run during sync() events.

Linux/Unix Note
A Linux GridLAB module must be a shared object library and must be listed in the appropriate makefile.am files.

Module implementation as of Hassayampa (Version 3.0)

The above implementation protocol is simplified using the C++ Module API.

main.cpp:

#define DLMAIN
#include <stdlib.h>
#include <stdio.h>
#include <errno.h>
#include "gridlabd.h"
#include "myclass.h"
EXPORT CLASS *init(CALLBACKS *fntable, MODULE *module, int argc, char *argv[])
{
	// set the GridLAB core API callback table
	callback = fntable;
 
	// TODO: register each object class by creating its first instance
	new myclass(module);

	// always return the first class registered
	return myclass::oclass;
}
CDECL int do_kill()
{
	// if anything needs to be deleted or freed, this is a good time to do it
	return 0;
}

myclass.h:

#ifndef _MYCLASS_H
#define _MYCLASS_H
#include "gridlabd.h"
class myclass : public gld_object {
public: // declare typedefs
  // declare properties using GL_ATOMIC, GL_STRING, GL_ARRAY, GL_STRUCT
private: // declare private members
public:
  myclass(MODULE *mod);
  // add declarations for create, init, presync, sync, postsync, precommit, commit, finalize, prenotify, postnotify, etc.
public:
  static CLASS *oclass;
  static myclass *defaults;
  // add pclass only if derived from another registered class
  static CLASS *pclass;
};
#endif // _MYCLASS_H

myclass.cpp when not derived from a registered class

#include "myclass.h"
EXPORT_CREATE(myclass)
// add EXPORT_* as needed (note that sync and notify are grouped)
myclass::myclass(MODULE *mod)
{
  if ( !oclass ) { // first time
    oclass = gld_class::create(mod,"myclass",sizeof(myclass),passconfig);
    if ( !oclass ) exception("unable to register class myclass");
    oclass->trl = TRL_UNKNOWN;
    defaults = this;
    if ( gl_public_variable(oclass,
         // add type,"name", get_name_offset(), PT_DESCRIPTION,"description",
         NULL)<1) { exception("unable to publish myclass properties"); }
    memset(this,0,sizeof(myclass));
    // add other default values as needed
  }
}
// add implementations for create, init, presync, sync, postsync, precommit, commit, finalize, prenotify, postnotify, etc.

myclass.cpp when derived from a registered class (with difference in bold)

#include "myclass.h"
EXPORT_CREATE(myclass)
// add EXPORT_* as needed (note that sync and notify are grouped)
myclass::myclass(MODULE *mod) : parentclass(mod)
{
  if ( !oclass ) { // first time
    pclass = parentclass::oclass;
    oclass = gld_class::create(mod,"myclass",sizeof(myclass),passconfig);
    if ( !oclass ) exception("unable to register class myclass");
    oclass->trl = TRL_UNKNOWN;
    defaults = this;
    if ( gl_public_variable(oclass,
         PT_INHERIT, "parentclass",
         // add type,"name", get_name_offset(), PT_DESCRIPTION,"description",
         NULL)<1) { exception("unable to publish myclass properties"); }
    memset(this,0,sizeof(myclass));
    // add other default values as needed
  }
}
// add implementations for create, init, presync, sync, postsync, precommit, commit, finalize, prenotify, postnotify, etc.

Upgrading modules to Hassayampa (Version 3.0)

As of Hassayampa (Version 3.0) modules must support multithreaded communication between classes. To upgrade a module created prior to Hassayampa (Version 3.0) full R/W locking of class members must be supported. The following must be done to enable use of locked accessors.

Minimum upgrade

The following steps are the minimum required to upgrade a class created prior to Hassayampa (Version 3.0) to full compatibility with Hassayampa (Version 3.0) and later:

  1. Change main.cpp to contain only the following
#define DLMAIN
#include "gridlabd.h"
  1. Derive all registered base classes from gld_object.
  2. Add PC_AUTOLOCK to the class registration pass configuration for any class that requires locking.
  3. Remove all self-locking calls.
  4. Add read or write lock/unlock calls around all accesses to other objects.
  5. Move first sync initialization code when t0=0 to init() and return deferred status when necessary (see Initialization Technical Manual).

Full upgrade

Update the main.cpp to contain only the following
#define DLMAIN
#include "gridlabd.h"
Derive existing base classes from gld_object
class myclass : public gld_object {
  // declarations ...
}
Change declarations of published variables to define get/set accessors
class myclass : protected gld_object {
  // declare members that can updated atomically (e.g., intN, double, enum, set)
  GL_ATOMIC(type,name);
  // declare string members (e.g., charN)
  GL_STRING(type,name);
  // declare struct members (e.g., struct, class)
  GL_STRUCT(type,name);
  // declare array members (e.g., anything[N])
  GL_ARRAY(type,name,size);
}
Replace all internal references to members with get/set accessors
Lookups (offsets)
PADDR(value)

become

get_value_offset()

and defaults=this; must be moved to before first use of get_*_offset().

Right-hand-side uses of value
atomic_value = get_atomic_var();
struct_value = get_struct_var();
char_ptr = get_string_var();
char_value = get_string_var(index);
array_ptr = get_array_var();
array_value = get_array_var(index);
Left-hand-side uses of value
set_atomic_var(atomic_value);
set_struct_var(struct_value);
set_string_var(char_ptr);
set_string_var(index,char_value);
set_array_var(array_ptr);
set_array_var(index,array_value);
Replace needed core linkage code with standard export macros
EXPORT_CREATE(my_class);
EXPORT_INIT(my_class);
EXPORT_COMMIT(my_class);
EXPORT_SYNC(my_class);
EXPORT_NOTIFY(my_class);
EXPORT_PRECOMMIT(my_class);
EXPORT_FINALIZE(my_class);
EXPORT_ISA(my_class);
EXPORT_PLC(my_class);
EXPORT_RECALC(my_class);
Depending on how the original export function was implemented, you may have to move

some code into the class member.

Move old-style initialization in sync() function to init(). The clock will no longer

be 0 on the first sync call. It will be set to the start time prior to init().

Surround all internal references requiring coherent access with locked blocks
rlock();
copy1 = var1;
copy2 = var2;
runlock();
wlock();
var1 = value1;
var2 = value2;
wunlock();
For programming convenience a scope lock is available
{ gld_wlock _lock(my); // write locked when block is entered
  // write safe code
  value1 += 12.3;
  value2 -= 12.3;
} // unlock when scope is exited
Ideally, enable automatic locking for all sync operations by added the PC_AUTOLOCK to the class registration call:
oclass = gl_register_class(module,"myclass",sizeof(myclass),passconfig|PC_AUTOLOCK);

Be careful not to use locking accessors within scope locks or autolocked sync functions. Doing so will cause a deadlock.

Modify all external references to use set/get accessors
gld_property prop(obj,"varname");
if ( prop.is_valid() )
{
   double value;
   prop.getp(value); // read lock is automatic for non-atomic operations
   value = 12.3;
   prop.setp(value); // write lock is automatic for non-atomic operations
}
If code requires a locked block use a scope lock w/non-locking accessors
gld_property prop(obj,"varname");
if ( prop.is_valid() )
{
   double value;
   gld_wlock lock(obj); // write lock is taken
   prop.getp(value,lock); // write lock continues so read is ok
   value += 12.3;
   prop.setp(value,lock); // write lock continues so write is ok
}  // unlocked

Debugging lock timeouts

A maximum spin timeout is implemented in both read and write locks to prevent deadlocks. If you run into a situation where you get a "write lock timeout" or "read lock timeout" then most likely you've encountered a condition where an object is trying to take a lock out on itself. Consider the following

Are you using a locking accessor in a sync call on an autolocked object?
If so, consider using direct access call instead because the object is already locked by the core.
Is the object locked for a very long time because it's doing something that really does take a long time?
If so, you may have to increase the value of MAXSPIN in the core/lock.cpp. Note that this value has a maximum of about 4e9 and it is currently set to 1e9 (which is roughly 10 seconds), so there's not much room left for growth. If this is a problem, please file a ticket and an alternate timeout method will have to be implemented.
Try debugging with a breakpoint on the throw_exception calls in core/lock.cpp
This should tell you exactly which lock is causing the timeout.

See also