On Thu, Jan 26, 2012 at 1:29 PM, Millies, Sebastian <[email protected]> wrote: > > From: Simon Laws [mailto:[email protected]] > Sent: Tuesday, January 24, 2012 5:08 PM > To: [email protected] > Subject: Re: Performance optimization of binding.sca ? > >>On Fri, Jan 20, 2012 at 4:01 PM, Millies, Sebastian >><[email protected]> wrote: >>>>From: Simon Laws [mailto:[email protected]] >>>>Sent: Friday, January 20, 2012 11:57 AM >>>>To: [email protected] >>>>Subject: Re: Performance optimization of binding.sca ? >>>> >>>>On Fri, Jan 20, 2012 at 8:10 AM, Millies, Sebastian >>>><[email protected]> wrote: >>>>>Hello there, >>>>> >>>>>I have a question about the performance optimization of binding.sca in >>>>>Tuscany 1.6. >>>>>Suppose I manually create and start two nodes in the same JVM, as follows: >>>>> >>>>>SCANodeFactory factory = SCANodeFactory.newInstance(); >>>>>SCANode node_1 = factory.createSCANode( composite_1, contributions_1 ); >>>>>SCANode node_2 = factory.createSCANode( composite_2, contributions_2 ); >>>>>node_1.start(); >>>>>node_2.start(); >>>>> >>>>>Suppose further that component A in composite 1 references a service of >>>>>component B in >>>>>composite 2, and both the reference and service expose <binding.sca/> . >>>>> >>>>>Will there be any internal emulation of a remote call, SOAP parsing etc? >>>>>Or will Tuscany make >>>>>a direct call, perhaps using Java reflection? >>>>> >>>>> -- Sebastian >>>>> >>>>Hi Sebastien. >>>> >>>>The idea was that an optimization would be used in that scenario where data >>>>is serialized across class loaders but without the requirement for a full >>>>remote protocol transformation. While we can detect the scenario I don't >>>>think the optimization is in place. >>>> >>>>Regards >>>> >>>>Simon >>> >>> Can you say something about HOW to detect that scenario and WHERE to apply >>> the optimization? >>> In due course I might have occasion to look at implementing that. >>> >> >>Apologies for the delayed response Sebastien. >> >>The place to start is the SCA Binding provider on the reference side. >>You can see code here that is deciding whether it thinks the target is >>remote or not. In the case that that target isn't remote we'd then >>need to think about the optimization. I.e. can we detect the situation >>where the target component is in the same VM but in a different >>classloader say. > > > It is important to see that in my example the target is not only in a > different classloader but also in a different node, defined by a different > composite. Classloader separation will also occur between different > contributions to the same node, but that's node the only thing going on here. > > >>For a target component that is in the same VM but has >>a remotable interface there is some code that enforces pass by value >>semantics (I think by serializing the arguments using the databinding >>layer). Have a look in the code where the PassByValueInterceptor [2] >>is set up >> >>Simon >> >>[1] >>http://svn.apache.org/repos/asf/tuscany/sca-java-1.x/trunk/modules/binding-sca/src/main/java/org/apache/tuscany/sca/binding/sca/impl/RuntimeSCAReferenceBindingProvider.java >>[2] >>http://svn.apache.org/repos/asf/tuscany/sca-java-1.x/trunk/modules/core-databinding/src/main/java/org/apache/tuscany/sca/core/databinding/wire/PassByValueInterceptor.java > > > This is getting more complicated than I thought it would be. I have > experimented a > bit, and now I do no longer even understand what is meant by a "local > interface". > > The SCA assembly spec says "A service typed by a local interface can only be > called by clients > that are running in the same process as the component that implements the > local service". > Something similar is said in Tuscany in Action: "Services with local > interfaces are to communicate > only with other components that are deployed locally and that are running > within a single JVM." > > The emphasis here is quite clearly on operating system processes, which are > equated with JVMs. > So in the code I posted originally (see above) it should be obvious that the > services in node1 > and node2 must be local with respect to each other, although defined in > different composites. > But this is not how Tuscany seems to work. > > For local services, I would expect to be able to use binding.sca with local > URIs. For > example, I would expect to be able to reference service "Callee" in node2 > from a > component in node1 using something like: > > <sca:reference name="callee" target="CalleeComponent/Callee"> > <sca:interface.java interface="caller.callee.api.Callee"/> > <sca:binding.sca/> > </sca:reference> > > where the service in node2 is simply exposed by: > > <sca:service name="Callee"> > <sca:interface.java interface="caller.callee.api.Callee"/> > <sca:binding.sca/> > </sca:service> > > I would further expect that the Callee-Interface may be non-remotable. > > On the contrary, Tuscany forces me to make Callee remotable, and specify > non-local targets and Uris for the reference and service, e. g. something like > http://localhost:8090/ps.softwareag.com/callee/Callee";, as if they were in > different processes. > > The code that checks whether the target is remote or not will always conclude > that it's remote if it has a remotable interface. > > Interestingly, the RMI-Binding is different: With the RMI-Binding I can > have a non-remotable interface. I will then get pass-by-reference as well, > i. e. no PassByValueInterceptor is inserted. (The RMI-Binding has a different > bug: When I make the interface remotable, any @AllowsPassByReference on the > Callee-Implementation is ignored, which is also not according to spec, but > not bothersome.) > > I believe the RMI approach will work for me. But I'd still be grateful for > some clarification. I could post illustrative code, if anyone is interested. > > -- Sebastian > > > IDS Scheer Consulting GmbH > Geschäftsführer/Managing Directors: Kamyar Niroumand, Ivo Totev > Sitz/Registered office: Altenkesseler Straße 17, 66115 Saarbrücken, Germany - > Registergericht/Commercial register: Saarbrücken HRB 19681 > http://www.softwareag.com >
I'm struggling a bit to remember how this works in 1.x (it's different in 2.x). There used to be code somewhere that detected this two nodes in the same JVM case but I can't put my finger on it. From your comments it's obviously not working. In 1.x we don't have a runtime registry where we can go an look up endpoints, as we do in 2.x, so we rely on the node processing and composite configuration to detect the different situations. Let me take a look at the code and see if I can shed any light on it. Simon -- Apache Tuscany committer: tuscany.apache.org Co-author of a book about Tuscany and SCA: tuscanyinaction.com
