This is a good optimization and should be considered the next time we devote
time to that.

Bug #358 describes a possible way to implement fragments: the CDR stream
classes can have a back-end that is used to allocate the buffers, when
an output CDR stream runs out of buffer space it can request more memory from
the back-end, but instead of performing another allocation the CDR back-end
sends a fragment, and continues marshaling on anew CDR stream.

A slightly clearer rewording of the tasks:

      . TAO allocates large buffers for applications sending large
        amounts of data, which increases the runtime memory of the
        application.  It would therefore be beneficial to reduce the
        buffer allocations by making use of the GIOP fragments feature.
        The goal is to perform CDR stream reads and writes on-demand,
        i.e. read/write enough to be able know what to demarshal or
        marshal, respectively.  The following link has details on how we
        propose to implement that in TAO:

        http://deuce.doc.wustl.edu/bugzilla/﷒0﷓
        http://deuce.doc.wustl.edu/bugzilla/﷒1﷓

        This task will involve the following activities:

        . Implement a "backing store" class that hides the source
          buffer, and provides on-demand reads and writes from and to
          that buffer, respectively.

        . Update the ACE CDR stream implementation and the TAO
          messaging code to utilize on-demand reads/writes.

Mine.

taking this over

accept

to pool