Serialization

Special Considerations

Object Tracking
Exporting Class Serialization
Class Information
Archive Portability Numerics Traits
Binary Archives
XML Archives
DLLS - Serialization and Runtime Linking
Multi-Threading
Optimizations
Archive Exceptions
Exception Safety

Object Tracking

Depending on how the class is used and other factors, serialized objects may be tracked by memory address. This prevents the same object from being written to or read from an archive multiple times. These stored addresses can also be used to delete objects created during a loading process that has been interrupted by throwing of an exception.

This could cause problems in progams where the copies of different objects are saved from the same address.


template<class Archive>
void save(boost::basic_oarchive  & ar, const unsigned int version) const
{
    for(int i = 0; i < 10; ++i){
        A x = a[i];
        ar << x;
    }
}

In this case, the data to be saved exists on the stack. Each iteration of the loop updates the value on the stack. So although the data changes each iteration, the address of the data doesn't. If a[i] is an array of objects being tracked by memory address, the library will skip storing objects after the first as it will be assumed that objects at the same address are really the same object.

To help detect such cases, output archive operators expect to be passed const reference arguments.

Given this, the above code will invoke a compile time assertion. The obvious fix in this example is to use


template<class Archive>
void save(boost::basic_oarchive & ar, const unsigned int version) const
{
    for(int i = 0; i < 10; ++i){
        ar << a[i];
    }
}

which will compile and run without problem. The usage of const by the output archive operators will ensure that the process of serialization doesn't change the state of the objects being serialized. An attempt to do this would constitute augmentation of the concept of saving of state with some sort of non-obvious side effect. This would almost surely be a mistake and a likely source of very subtle bugs.

Unfortunately, implementation issues currently prevent the detection of this kind of error when the data item is wrapped as a name-value pair.

A similar problem can occur when different objects are loaded to and address which is different from the final location:


template<class Archive>
void load(boost::basic_oarchive  & ar, const unsigned int version) const
{
    for(int i = 0; i < 10; ++i){
        A x;
        ar >> x;
        std::m_set.insert(x);
    }
}

In this case, the address of x is the one that is tracked rather than the address of the new item added to the set. Left unaddressed this will break the features that depend on tracking such as loading object through a pointer. Subtle bugs will be introduced into the program. This can be addressed by altering the above code thusly:


template<class Archive>
void load(boost::basic_iarchive  & ar, const unsigned int version) const
{
    for(int i = 0; i < 10; ++i){
        A x;
        ar >> x;
        std::pair<std::set::const_iterator, bool> result;
        result = std::m_set.insert(x);
        ar.reset_object_address(& (*result.first), &x);
    }
}

This will adjust the tracking information to reflect the final resting place of the moved variable and thereby rectify the above problem.

If it is known a priori that no pointer values are duplicated, overhead associated with object tracking can be eliminated by setting the object tracking class serialization trait appropriately.

By default, data types designated primitive by Implementation Level class serialization trait are never tracked. If it is desired to track a shared primitive object through a pointer (e.g. a long used as a reference count), It should be wrapped in a class/struct so that it is an identifiable type. The alternative of changing the implementation level of a long would affect all longs serialized in the whole program - probably not what one would intend.

It is possible that we may want to track addresses even though the object is never serialized through a pointer. For example, a virtual base class need be saved/loaded only once. By setting this serialization trait to track_always, we can suppress redundant save/load operations.


BOOST_CLASS_TRACKING(my_virtual_base_class, boost::serialization::track_always)

Exporting Class Serialization

Elsewhere in this manual, we have described BOOST_CLASS_EXPORT. This is used to make the serialization library aware that code should be instantiated for serialization of a given class even though the class hasn't been otherwise referred to by the program. This functionality is necessary to implement serialization of pointers through a virtual base class pointer. That is, a polymorphic pointer.

This macro specifies a "Globally Unique IDentifier". This is an string which identifies the class to be created when data is loaded. Generally a text representation of the class name is sufficient for this purpose, but in certain cases it maybe necessary to specify a different string by using BOOST_CLASS_EXPORT_GUID rather than a simple BOOST_CLASS_EXPORT.

BOOST_CLASS_EXPORT would usually be specified in the same header file as the class declaration to which it corresponds. That is, BOOST_CLASS_EXPORT(T) is a "trait" of the class T. So a program using this class will look something like:


#include <boost/archive/xml_oarchive.hpp>
.... // any other archive classes
#include "my_class.hpp"  // which contains BOOST_CLASS_EXPORT(my_class)

These headers can be in any order. (In boost versions 1.34 and earlier, the archive headers had to go before any headers which contain BOOST_CLASS_EXPORT.) Any code required to serialize types specified by BOOST_CLASS_EXPORT will be instantiated for each archive whose header is included. (note that the code is instantiated regardless of whether or not it is actually invoked.) If no archive headers are included - no code should be instantiated. This will permit BOOST_CLASS_EXPORT to be a permanent part of the my_class.hpp .

Note that the implementation of this functionality depends upon vendor specific extensions to the C++ language. So, there is no guarenteed portability of programs which use this facility. However, all C++ compilers which are tested with boost provide the required extensions. The library includes the extra declarations required by each of these compilers. It's reasonable to expect that future C++ compilers will support these extensions or something equivalent.

Class Information

By default, for each class serialized, class information is written to the archive. This information includes version number, implementation level and tracking behavior. This is necessary so that the archive can be correctly deserialized even if a subsequent version of the program changes some of the current trait values for a class. The space overhead for this data is minimal. There is a little bit of runtime overhead since each class has to be checked to see if it has already had its class information included in the archive. In some cases, even this might be considered too much. This extra overhead can be eliminated by setting the implementation level class trait to: boost::serialization::object_serializable.

Turning off tracking and class information serialization will result in pure template inline code that in principle could be optimised down to a simple stream write/read. Elimination of all serialization overhead in this manner comes at a cost. Once archives are released to users, the class serialization traits cannot be changed without invalidating the old archives. Including the class information in the archive assures us that they will be readable in the future even if the class definition is revised. A light weight structure such as display pixel might be declared in a header like this:


#include <boost/serialization/serialization.hpp>
#include <boost/serialization/level.hpp>
#include <boost/serialization/tracking.hpp>

// a pixel is a light weight struct which is used in great numbers.
struct pixel
{
    unsigned char red, green, blue;
    template<class Archive>
    void serialize(Archive & ar, const unsigned int /* version */){
        ar << red << green << blue;
    }
};

// elminate serialization overhead at the cost of
// never being able to increase the version.
BOOST_CLASS_IMPLEMENTATION(pixel, boost::serialization::object_serializable);

// eliminate object tracking (even if serialized through a pointer)
// at the risk of a programming error creating duplicate objects.
BOOST_CLASS_TRACKING(pixel, boost::serialization::track_never)

Archive Portability

Several archive classes create their data in the form of text or portable a binary format. It should be possible to save such an of such a class on one platform and load it on another. This is subject to a couple of conditions.

Numerics

The architecture of the machine reading the archive must be able hold the data saved. For example, the gcc compiler reserves 4 bytes to store a variable of type wchar_t while other compilers reserve only 2 bytes. So its possible that a value could be written that couldn't be represented by the loading program. This is a fairly obvious situation and easily handled by using the numeric types in <boost/cstdint.hpp>

A special integral type is std::size_t which is a typedef of an integral types guaranteed to be large enough to hold the size of any collection, but its actual size can differ depending on the platform. The collection_size_type wrapper exists to enable a portable serialization of collection sizes by an archive. Recommended choices for a portable serialization of collection sizes are to use either 64-bit or variable length integer representation.

Traits

Another potential problem is illustrated by the following example:


template<class T>
struct my_wrapper {
    template<class Archive>
    Archive & serialize ...
};

...

class my_class {
    wchar_t a;
    short unsigned b;
    template<<class Archive>
    Archive & serialize(Archive & ar, unsigned int version){
        ar & my_wrapper(a);
        ar & my_wrapper(b);
    }
};

If my_wrapper uses default serialization traits there could be a problem. With the default traits, each time a new type is added to the archive, bookkeeping information is added. So in this example, the archive would include such bookkeeping information for my_wrapper<wchar_t> and for my_wrapper<short_unsigned>. Or would it? What about compilers that treat wchar_t as a synonym for unsigned short? In this case there is only one distinct type - not two. If archives are passed between programs with compilers that differ in their treatment of wchar_t the load operation will fail in a catastrophic way.

One remedy for this is to assign serialization traits to the template my_template such that class information for instantiations of this template is never serialized. This process is described above and has been used for Name-Value Pairs. Wrappers would typically be assigned such traits.

Another way to avoid this problem is to assign serialization traits to all specializations of the template my_wrapper for all primitive types so that class information is never saved. This is what has been done for our implementation of serializations for STL collections.

Binary Archives

Standard stream i/o on some systems will expand linefeed characters to carriage-return/linefeed on output. This creates a problem for binary archives. The easiest way to handle this is to open streams for binary archives in "binary mode" by using the flag ios::binary. If this is not done, the archive generated will be unreadable.

Unfortunately, no way has been found to detect this error before loading the archive. Debug builds will assert when this is detected so that may be helpful in catching this error.

XML Archives

XML archives present a somewhat special case. XML format has a nested structure that maps well to the "recursive class member visitor" pattern used by the serialization system. However, XML differs from other formats in that it requires a name for each data member. Our goal is to add this information to the class serialization specification while still permiting the the serialization code to be used with any archive. This is achived by requiring that all data serialized to an XML archive be serialized as a name-value pair. The first member is the name to be used as the XML tag for the data item while the second is a reference to the data item itself. Any attempt to serialize data not wrapped in a in a name-value pair will be trapped at compile time. The system is implemented in such a way that for other archive classes, just the value portion of the data is serialized. The name portion is discarded during compilation. So by always using name-value pairs, it will be guarenteed that all data can be serialized to all archive classes with maximum efficiency.

DLLS - Serialization and Runtime Linking

Serialization code can be placed in libraries to be linked at runtime. That is, code can be placed in DLLS(Windows) or Shared Libraries(*nix). Along with the "export" facility, this permits a program to written without knowledge of the actual types to be serialized. This package doesn't include an example of this technique - but coding would be very similar to the example demo_pimpl.cpp , demo_pimpl_A.cpp and demo_pimpl_A.hpp where implementation of serializaton is completely separate from the main program.

Multi-Threading

The nature of serialization would conflict with multiple thread concurrently writing/reading from/to a single open archive. Since each archive is independent from ever other one, there should be no problem in having multiple open archives from one or more threads.

Well, not quite.

There are a couple of global data structures for holding information of serializable types. These structures are used to dispatch to correct code to handle each pair of serializable types and archive types. Since this information is shared among all archives, there is potential for problems. This has been addressed carefully implementing the library so that these structures are all initialized before main(...) is called. From then on they are never altered. So there SHOULD be no problem having mulitple archives open simultaneously - be it from the same or different threads.

Well, almost.

With dynamically loaded code - DLLS or Shared Libraries, these global data structures can be altered when a library is loaded or unloaded. That is, in this case, these globa data structures can be altered after main(...) is called. So if a thread is dynamically loading/unloading modules which contain serialization code while an archive is open there could be problems. Also, if such loading/unloading is happening concurrently in different threads, there could also be problems.

It might not be easy to control this. Is possible that some systems may not actually load modules until they are actually needed. So even though we think that there is not dynamic loading/unloading of such code it could be occurring as "help" to manage resources. On such systems, access to archive code would have to be syncronized with some multi-threading construct in order to be functional.

Optimizations

In performance critical applications that serialize large sets of contiguous data of homogeneous types one wants to avoid the overhead of serializing each element individually, which is the motivation for the array wrapper. Serialization functions for data types containing contiguous arrays of homogeneous types, such as for std::vector, std::valarray or boost::multiarray should serialize them using an array wrapper to make use of these optimizations. Archive types that can provide optimized serialization for contiguous arrays of homogeneous types should implement these by overloading the serialization of the array wrapper, as is done for the binary archives.