Dennis Noll had a very good suggestion to optimize the critical path in TAO:

> TSS CACHED OUTPUT CDRs - Instead of creating an Output CDR & buffer
> (4096 bytes) on the stack for each invocation, I changed the output
> CDR attribute in IIOP_Invocation to be a reference which the
> constructor retrieves from TSS.  This saves one TSS lookup (2
> allocators vs 1 lookup for the entire CDR) and a lot of layered
> constructor/destructor code (TAO_OutputCDR, ACE_OutputCDR, ACE_CDR,
> ACE_Message_Block, ACE_Timer, etc., etc., etc.).  Same logic applied
> to the server side, where an output CDR was created for all
> operations whether they need it or not (eg - one-ways).

Unfortunately his idea does not apply for all applications, in particular,
if you use nested upcalls you cannot expect that a single TSS CDR stream
will be usable.
But i think there is a way to generalize his idea: we could keep a TSS pool
of OutputCDR streams, the thread will pick one from the pool and return it
when the GIOP message is sent.

I thought about allocating the OutputCDR stream on the stack of the
run() or perform_work() method, but i don't think that will improve
anything.