RMI optimizations (stub sharing, local calls)

Question

Hi,
Suppose I have the following scenario involving one client and one server:
1. the server exports a RMI object implementing "MyInterface" with a method "void op(MyInterface myObject)"
2. the client obtains two references to the same remote object instance, lets call them myObjectRef1 and myObjectRef2
3. the client calls the given method with one of the obtained references as a parameter, say myObjectRef1.op(myObjectRef2);
My two questions are:
In step 2: Will the RMI implementation automatically recognize that both remote reference point to the same object instance and use one shared stub? Or will there be two totally independent stubs?
In step 3: When the client invokes the method, the stub that is passed points to a RMI object that could be accessed via a local object reference on the server that is executing the method.
Theroretically either the serialized stub could be substituted by a local object reference when the method call is unmarshalled or at least upon invoking a method on the stub, the RMI system could optimize away marshalling and TCP/IP by directly calling the implementation method in the RemoteRef.invoke method. Is there some optimization like this in the Sun RMI implementation? Or is it totally unaware of such scenarios and just performs a remote call involving marshalling+TCP/IP communication?
Thanks for your replys.

Answer 1

In step 2: Will the RMI implementation automatically recognize that both remote reference point to the same object instance and use one shared stub? Or will there be two totally independent stubs?As a matter of fact there is a shared stub at the remote object, but by the time they're acquired by the client they will be distinct, because of serialization over distinct ObjectOutputStreams (a new one per call). But why would it matter to you either way? How could you even tell? The stubs all point to the same remote object, and the equals() methods return true.
In step 3: When the client invokes the method, the stub that is passed points to a RMI object that could be accessed via a local object reference on the server that is executing the method.If it's a stub it will always perform RMI.
Theroretically either the serialized stub could be substituted by a local object reference when the method call is unmarshalled or at least upon invoking a method on the stub, the RMI system could optimize away marshalling and TCP/IP by directly calling the implementation method in the RemoteRef.invoke method.No, it can't. That would change the pass-by-copy semantics of RMI into pass-by-value as in local method invocation. It can't do that.
Is there some optimization like this in the Sun RMI implementation?No. Nor in IBM's either.