Visibroker 5.2.1 callbacks optimization - Object

ivandric@yahoo.co.uk wrote in message news:<5a3f642c.0409140628.30dee9c1@posting.google.com>...
> Hi
>
> I have a fairly simple Corba server application
> that receives the data in a form of a callback
> and republishes that data further.
> The callback should consume incoming data as soon as possible.
> Visibroker allocates thread for every callback.
>
> There are around 300 callback Corba objects and Visibroker
> runs multiple threads on every object.
> I have configured the Visibroker to have a thread pool with 1000 threads.
>
> The code looks something like this:
>
> CORBA::ORB_var orb = CORBA::ORB_init(argc, argv);
> CORBA::Object_var obj = orb->resolve_initial_references("RootPOA");
> PortableServer::POA_var rootPOA = PortableServer::POA::_narrow(obj);
> PortableServer::POAManager_var poa_mgr = rootPOA->the_POAManager();
>
> vector< MyCallback* > myCallback_vec;
> for( int i = 0; i < symbols.size(); ++i)
> {
> myCallback* p_c = new MyCallback;
> rootPOA->activate_object( p_c);
> myCallback_vec.push_back( p_c);
> }
>
> poa_mgr->activate();
>
> // logic for callback subscriptions
>
> orb->run();
>
> I would very much appreciate advice on how to make this faster.
> What can I do to have callbacks cosume data faster?
> Perhaps, create additional POA's or something else?
>

Setting thread max to 1000 is too high and likely harmful for
performance. Unless the servant operation has high blocking rate (for
instance, operation blocks on remote call to a database, etc. 80% of
the time), large number of concurrent, CPU consumed tasks will only
degrade the performance by unnecessary high rate of context switch.
Therefore, if your servant operation's implementation is non-blocking
and cpu consumed, I would suggest setting thread pool size not too
much higher than the number of available CPU on the host.

The second suggestion is to use a notification service for this kind
of subscribe/publish architecture. A decent notification service (e.g.
Borland VisiNotify) should support typed event and also should have
much better performance (e.g. throughput is 3 to 10 times faster) than
homegrown solutions.

Regards,
Ke

> Regards,
> Ivo