Saving and Restoring States of CLHEP Random Engines and Distributions

We have added new (as of late 2004) methods to the CLHEP Random classes allowing more convenient and natural saving and restoring of engines and distributions. Some inconvenient features of the original methods are:
The saveStatus and restoreStatus methods are based on a named file, so that each distribution saved must be placed in its own file.
Since opening the file is left to the restoreStatusmethod, there can be no user-controlled flexibility in how to deal with trouble (such as file not found) in opening that file.
Each distribution saves the state of the engine it is using. If several distributions share a single instance of an engine, then this wastes I/O time and space on multiple copies of the same data.
There was never a way to save an engine, and then have another program instantiate a HepRandomEngine* and read back the state to create an engine of the appropriate type, without knowing what sort of engine was saved. We call this functionallity an "engine factory."
For cases where the user program has its own scheme to persist numbers (for example, unsigned longs), there is no way to cause an engine to produce its state (or restore from a saved state) as a vector of unsigned longs.


The usual idiom for saving data would be to write to an ostream and read
from an istream.  
It is assumed that users know how to construct an fstream
(ofstream or ifstream) to work with a file, and 
using these 
fstreams as ostreams and istreams.

We have added methods that stick to this idiom, which 
greatly helps matters.  (In fact, 
the first two of these inconveniences are immediately resolved, 
and the structure is right for dealing with the third.)

We have also added engine factory 
capability to deal with anonymous engine saves,
and we have given every engine the ability to save and restore to/from
a vector of unsigned longs.


  Use Cases 
  Added Methods 
  How the Use Cases Can Be Coded



 Use Cases 

 

Each of these motivating use cases are based on the following theme:
An experiment wishes, when running a simulation job in "debuggable" mode, 
to capture before each event enough information that a subsequent job can 
restart at an arbitrary critical event and get the same results (or
more typically, exhibit the same problem for investigation).
The states of all random engines evolve across events, so this state information
must be saved with each event (and restored for the critical event).



There have been occasional requests for methods to let a job
restore a "generic" engine (that is, to start using an engine
without knowing at compile time which type of engine will have been saved).
Some of these added methods provide that capability in special circumstances.
In general, however, 
we make the assumption that the restoring user knows the classes which have
been saved.  If the restore does not match the save, this will be indicated
in the usual ways; in the Random package, this includes output to 
cerr 
and setting the state of the restoring istream to "bad."


 One static distribution 

The user employs one static distribution, calling (say) 
RandGauss::shoot().
She wishes to save/restore the full state (including underlying engine)
of just that distribution. 
This is completely analogous to doing 
RandGauss::saveState("filename"),
but it is desired that this state be added to a file rather than
placed in its own file.

 One instance of a distribution 

The user has an instance of a distribution, which uses some instance 
of an engine.  She wishes to save/restore its full state, including that of 
the underlying engine.

 Some distributions sharing various engines 

The user has several instances of a distributions of various types, 
(and/or uses the shoot() 
method of several static distributions)
and one or more instances of engines of various types.  
To conserve I/O space, she wishes to separately save each engine and
the non-engine-dependant state of each distribution.
Since she knows which engines are used by which distributions, she will 
be able to restore the full states of all the distributions after restoring
individual engine and distribution states.  

 All the static distributions 

A multi-module program written by a collaboration makes extensive use
of the shoot() 
method of various static distributions.  
They wish to make a single call to save/restore the states of all the
static distributions in the CLHEP Randoms package.  
The program restoring the states of the distributions and
the common underlying engine has no a priori 
knowledge of which type of the underlying engine has been set.  

If the static distributions each used a separate engine, 
this method would need to
ensure that states of shared engines would be saved only once,
and that all needed engine instances would be saved.  
However, as currently implemented, every static distribution shares
one common engine.  
Since the (non-engine) state of a distribution itself is miniscule,
and only a couple of distributions contain state information other than
non-standard distribution parameters (which do not apply to static 
distributions), 
the "wasted" cost of saving any unused static distributions is acceptable.


 One engine of unknown type 

An application wishes to deal with an engine restored from a saved 
engine state, without compile-time
knowledge of which engine type this will be.
For example, the saved engine may be a product of a simulation job 
where the engine created was determined by run-time input.

 Saving and restoring state via a vector of unsigned longs 

An experiment framework has a means of imbedding an array of numbers into 
the data for an event, and would wish to use this to hold the state of
the engine(s) used.  Here, the model of saving to a stream is less convenient
and much less efficient.  HepMC assumes random engines can produce their
states in the form of a vector of unsigned longs.
  

 Added Methods in CLHEP Random 

 

There are two types of save and restore activities on a distribution,
just as there are two modes of using a distribution to generate a random 
variate, depending on whether you want the static distribution or an instance
of the distribution.  
For the generation activity, the methods are 
Class::shoot() 
to use the static distribution, and 
instance.fire() to use
an instance.  

For these added methods which deal with saving and restoring to 
an fstream, the static methods all take the stream as an argument, 
while the
instance methods use the 
ostream <<  whatever 
or
istream >>   whatever 
syntax.

As in the case of variate generation, all the methods dealing with just an 
engine are based on an instance.

 Static methods in the distribution classes 

Each distribution has six new static methods, 
which override static methods of the same names in the 
HepRandom
class:

  
static ostream& saveFullState   (ostream& os) const;
static istream& restoreFullState(istream& is);



Saves (or restores) the state of the distribution,
including the engine used when shoot() 
is invoked and any
additional cached data, such as the second gaussian in the Box-Mueller
method used by RandGauss.


  
static ostream& saveDistState   (ostream& os) const;
static istream& restoreDistState(istream& is);



Saves (or restores) the static state of the distribution,
including cached data used in the course of the 
shoot() method, 
so that if the engine used is also restored,
the full state will have been replicated.
For distributions which do not distributions contain cached data
(all distributions other than RandGauss, RandFlat, and RandBit) this method
has no effect.


  
static ostream& saveStaticRandomStates (ostream& os) const;



Determines the identity of  
theEngine 
used in the static shoot()
method of the distributions.
Saves the state of that engine, followed by the static state of each 
distribution, such that the states
can be restored by    
restoreStaticRandomStates.


  
static istream& restoreStaticRandomStates(istream& is);



Restores the cached data for every static distribution and 
the engine used as theEngine 
in the static shoot() 
method of the distributions.

If, at the point of calling 
restoreStaticRandomStates,
the common engine has been set to an engine of the type matching
that saved, then that engine's state will be changed to the saved state.

If the two engine types do not match, a new engine of the saved type will
be instantiated, and attached to the static distribuitions via a 
setTheEngine() call.  
The previously existing engine
will not be deleted (since it might be shared by some distribution 
objects), and the new engine is a post-startup new object that 
won't automatically be deleted, so this may be detected by tools such
as Purify as a memory leak.





 Instance methods in the distribution classes  

  
ostream& operator << (ostream& os, const HepRandom & distribution);
istream& operator >> (istream& is,       HepRandom & distribution);



Saves (or restores) the state of this instance of the distribution,
including cached data used in the course of the 
fire() method, 
so that if the engine used is also restored,
the full state will have been replicated.

  
  
RandomEngine * HepRandom::engine();



Provides a pointer to engine underlying this instance of distribution.
This is needed to save the full state of this instance of distribution.

  
Conventionally, when saving the full state of a distribution, one should
follow the order used by the methods for static distributions, which save the
engine and then the remainder of the distribution state.
 
 Instance methods in the engine classes  

Some added methods of engine are provided, as tools for implementation 
of "generic engine" capability:
  
string RandomEngine::name();



Provides a string identifying the engine type.

  
istream& RandomEngine::getName(istream&);



Reads a string identifying the engine type, and verifies that it is the 
expected identification.

  
istream& RandomEngine::getState(istream&);



Inputs engine state.

  
ostream& RandomEngine::put(ostream&);
istream& RandomEngine::get(istream&);



Outputs or inputs engine state, including identifying name.  

Used by operator<< and operator>>.

  
vector < unsigned long > RandomEngine::put() const;
bool RandomEngine::get(const vector &);



Writes or sets engine state, including 32-bit identifier, 
to a vector of unsigned longs.  


Methods for I/O of engine states to streams are already present.  
Though not "additional methods," this are listed here because they
are logically on the same footing as the distribution I/O methods, 
and will appear in the section describing how to code the use cases.

  
ostream& operator << (ostream& os, const RandomEngine & engine);
istream& operator >> (istream& is,       RandomEngine & engine);



Saves (or restores) the state of the engine, including an identifying string
to help catch mismatches where erroneous code tries to use an engine state 
as the state of a different kind of engine.


Finally, the base class HepRandomEngine 
has a pair of methods to obtain a pointer to a new engine based on 
"generic" or anonymous engine input:  

  
static HepRandomEngine* HepRandomEngine::newEngine(istream& is);
static HepRandomEngine* HepRandomEngine::newEngine
				(const vector & v);



Based on the contents of the istream or vector, 
which must contain the output of
saving some engine, this instantiates a new engine of the
saved type, and returns a pointer to that object.  The semantics are
that of new:  It is the user's responsibility to invoke delete 
when the engine is no longer needed.


Note that although the engine type is deduced from the input, if the
saving program had any non-trivial engine sharing among multiple distribution
objects, the restoring program would have to set up the same sharing to get
the same subsequent random sequence behaviors.



 Caveat concerning RandEngine 

In the case of an engine based on system random generator engine whose 
internal state is not controllable by the
Random package, such as 
RandEngine,
the additional save/restore capability is still presented. 
However, in such a case, the only way to achieve the restore is to 
start in the original seed state, and fire off however many 
variates were requested prior to the save.  In the case of restoring
state based on a late event in a long job, this may be quite 
inefficient.  Therefore, we heavily recommend against using 
RandEngine in applications 
involving saving and restoring engine states.


 How the Use Cases Can Be Coded 

 

 One static distribution 

The "saving job" does:

 
  DualRand e(24681357);         // Illustrating that the engine need not 
  RandGauss::setTheEngine (&e); // be the default engine for RandGauss.
  ofstream fs(filename);        // Open it, or pass an ofstream already in use
  codeUsingRandGaussManyTimes();
 
  RandGauss::saveFullState(fs);




The "restoring job" does:

 
  DualRand e();   
  RandGauss::setTheEngine (e);  
  ifstream file(filename);      // Open it, or pass an ifstream already in use
 
  RandGauss::restoreFullState(fs);  





 One instance of a distribution 

The "saving job" does:

 
  DualRand e(24681357);   
  RandGauss g(e);               // Normal distribution, based on engine e  
  ofstream fs(filename);        
  codeUsing_g_ManyTimes();
 
  fs << g.engine() << g;




The "restoring job" does:

 
  DualRand e();   
  RandGauss g(e);
  ifstream fs(filename);      
 
  fs >> g.engine() >> g; 




 Some distributions sharing various engines 

The "saving job" does:

 
  DualRand e1(24681357);  
  MTwistEngine e2(135797531); 
  RandGauss g(e1);
  RandFlat  f(e1); 
  RandBit   b(e2);                
  ofstream fs(filename);        
  codeUsing_g_f_and_b();
 
  fs << e1 << e2 << g << f << b;




The "restoring job" does:

 
  e = new DualRand();   
  RandGauss g(e);
  ifstream fs(filename);      
 
  fs >> e1 >> e2 >> g >> f >> b;




To save and restore successfully in the general case, 
you must know not only which types of 
engines underlie the various distributions, but also which distributions 
share an engine.

 All the static distributions 

The "saving job" does:

 
  e = new DualRand(24681357);   // illustrating that the some distributions 
  RandGauss::setTheEngine (e);  // need not use the default engine
  ofstream fs(filename);        
  codeUsingVariousDistributions();
 
  saveStaticRandomStates(fs);




The "restoring job" does:

 
  e = new DualRand();   
  RandGauss::setTheEngine (e);  
  ifstream file(filename);      
 
  restoreStaticRandomStates(fs);




 One engine of unknown type 

The "saving job" does:

 
  DualRand *     e1 = new DualRand(24681357);
  RanecuEngine * e2 = new RanecuEngine(24563291);  
  HepRandomEngine *eptr;
  if (someControl) eptr = e1; else eptr = e2;  
  RandGauss g(*eptr);             
  ofstream fs(filename);        
  codeUsing_g_ManyTimes();
 
  fs << g.engine() << g;

  delete e1; delete e2; // (when g goes out of scope)



The "restoring job" does:

 
  ifstream fs(filename);      
 
  HepRandomEngine * eptr = HepRandomEngine::newEngine(fs);
  if (eptr==0) throw "problem inputting an engine";
  RandGauss g(*eptr);

  codeUsing_g();  // using g without knowing or caring 
                  // what type the engine is
  delete eptr;    // (delete engine when g goes out of scope) 



 Saving and restoring state via a vector of unsigned longs 
  
The "saving job" does:
 
 
  HepJamesRandom e = new HepJamesRandom(12343211);  
  codeUsing_e (e);
  std::vector < unsigned long > v;
 
  v = e.put();

  eventData.add_a_vector (v);



The "restoring job" does:

 
  eventData.add_a_vector (v);
  HepJamesRandom e;
 
  bool ok = e.get(v);
  if (!ok) throw "problem restoring an engine";

  codeUsing_e(e);  











Author: Mark Fischler


Last modified: March 15, 2005