Re: [boost] Serialization Draft Submission # 2

From: Vahan Margaryan (vahan_at_[hidden])
Date: 2002-03-22 00:53:50

> Fellow programmers
>
> I have just uploaded draft # 2 of my proposal for a boost library to
> address saving and restoring of C++ objects. Serialization.zip
>

I have written a similar system for my former company. Some of the things
that I have discovered that might be interesting to you in your development:

1. When loading containers, it's better to use their specific features to
achieve maximum performance. Specifically:

a) You can call 'reserve' for a vector, before reading in the elements.

b) When reading in map, set, multimap, multiset, you can call insert(end( ),
t) instead of insert(t). When you write out a tree-based associative
container, the elements are written out sorted. Therefore, when you read
them back, you can assume that each consecutive item should be inserted at
the end. The assumption isn't always valid, because sometimes the predicate
depends on volatile stuff, like values of pointers. In that case, insert at
end( ) is the same as plain insert. In the more frequent case, when
predicates are not sensitive to changes in pointer values, insert at end( )
is much faster.

c) Hash containers differ across implementations. Whenever the
implementation allows, it makes sense to determine an optimal number of hash
buckets during loading and call resize( ). Then both loading of the hash
table and the further operation of the container is faster.

These modifications affect the speed of loading very significantly, when
many objects and containers are deserialized. I have observed a 3x increase
in performance in one particular case.

2. Compatibility issues:

I tried to use the library under MSVC 6.0, and it worked, after a couple of
minor changes.
a) There is no <hash_map> in the standard library, and some libraries,
including the one coming with MSVC 6.0, don't have it in their distribution.
I commented the line including <hash_map> out in serialization.hpp. I guess
some kind of ifdef is needed there.

b) MSVC gives several compile errors related to comparison set::iterator ==
set::const_iterator. When I swap them (const_iterator == iterator), it
works. This is a bug in the old Dinkum library. The new Dinkum library
doesn't have this bug, but shipped with MSVC is the old one... Specific
locations where I get this error are:
serialization.hpp, line 930,
serialization.cpp, line 53
serialization.cpp, line 115

Additional compilation problems may arise, as more stuff gets
instantiated...

Regards,
-Vahan