Table of Contents
To minimize the time you have to wait on the compiler, it's good to
      only include the headers you really need.  Many people simply include
      <iostream> when they don't
      need to -- and that can penalize your runtime as well.
      Here are some tips on which header to use
      for which situations, starting with the simplest.
   
<iosfwd>
      should be included whenever you simply need the name
      of an I/O-related class, such as "ofstream" or
      "basic_streambuf".
      Like the name implies, these are forward declarations.
      (A word to all you fellow old school programmers:
      trying to forward declare classes like "class istream;"
      won't work.
      Look in the <iosfwd> header
      if you'd like to know why.)  For example,
   
    #include <iosfwd>
    class MyClass
    {
	....
	std::ifstream&   input_file;
    };
    extern std::ostream& operator<< (std::ostream&, MyClass&);
   <ios>
      declares the base classes for the entire I/O stream hierarchy,
      std::ios_base and std::basic_ios<charT>,
      the counting types std::streamoff and std::streamsize,
      the file positioning type std::fpos,
      and the various manipulators like std::hex,
      std::fixed, std::noshowbase,
      and so forth.
   
The ios_base class is what holds the format
      flags, the state flags, and the functions which change them
      (setf(), width(),
      precision(), etc).
      You can also store extra data and register callback functions
      through ios_base, but that has been historically
      underused.  Anything
      which doesn't depend on the type of characters stored is consolidated
      here.
   
The class template basic_ios is the highest
      class template in the
      hierarchy; it is the first one depending on the character type, and
      holds all general state associated with that type:  the pointer to the
      polymorphic stream buffer, the facet information, etc.
   
<streambuf>
      declares the class template basic_streambuf, and
      two standard instantiations, streambuf and
      wstreambuf.  If you need to work with the vastly useful and
      capable stream buffer classes, e.g., to create a new form of storage
      transport, this header is the one to include.
   
<istream>
       and <ostream>
       are the headers to include when you are using the overloaded
      >> and << operators,
      or any of the other abstract stream formatting functions.
      For example,
   
    #include <istream>
    std::ostream& operator<< (std::ostream& os, MyClass& c)
    {
       return os << c.data1() << c.data2();
    }
   The std::istream and std::ostream classes are the abstract parents of the various concrete implementations. If you are only using the interfaces, then you only need to use the appropriate interface header.
<iomanip>
      provides "extractors and inserters that alter information maintained by
      class ios_base and its derived classes,"
      such as std::setprecision and
      std::setw.  If you need
      to write expressions like os << setw(3); or
      is >> setbase(8);, you must include
      <iomanip>.
   
<sstream>
      and <fstream>
      declare the six stringstream and fstream classes.  As they are the
      standard concrete descendants of istream and ostream,
      you will already know about them.
   
Finally, <iostream>
      provides the eight standard global objects
      (cin, cout, etc).  To do this correctly, this
      header also provides the contents of the
      <istream> and
      <ostream>
      headers, but nothing else.  The contents of this header look like:
   
    #include <ostream>
    #include <istream>
    namespace std
    {
	extern istream cin;
	extern ostream cout;
	....
	// this is explained below
	static ios_base::Init __foo;    // not its real name
    }
   Now, the runtime penalty mentioned previously:  the global objects
      must be initialized before any of your own code uses them; this is
      guaranteed by the standard.  Like any other global object, they must
      be initialized once and only once.  This is typically done with a
      construct like the one above, and the nested class
      ios_base::Init is
      specified in the standard for just this reason.
   
How does it work? Because the header is included before any of your code, the __foo object is constructed before any of your objects. (Global objects are built in the order in which they are declared, and destroyed in reverse order.) The first time the constructor runs, the eight stream objects are set up.
The static keyword means that each object file compiled
      from a source file containing
      <iostream> will have its own
      private copy of __foo.  There is no specified order
      of construction across object files (it's one of those pesky NP complete
      problems that make life so interesting), so one copy in each object
      file means that the stream objects are guaranteed to be set up before
      any of your code which uses them could run, thereby meeting the
      requirements of the standard.
   
The penalty, of course, is that after the first copy of __foo is constructed, all the others are just wasted processor time. The time spent is merely for an increment-and-test inside a function call, but over several dozen or hundreds of object files, that time can add up. (It's not in a tight loop, either.)
The lesson?  Only include
      <iostream> when you need
      to use one of
      the standard objects in that source file; you'll pay less startup
      time.  Only include the header files you need to in general; your
      compile times will go down when there's less parsing work to do.