Networked Scene Graph (NSG)


The task is to come up with a mechanism that allows existing applications to run across a cluster of workstations. Each node on the cluster runs its own copy of the application and data, rendering only a specific view of the application. One of the nodes in designated the master and the rest are slaves. The master handles any user-interaction and propagates its scene graph changes to the slaves through the CAVERN dbserver.

The NSG API defines the protocol for this uni-directional communication. The core of the API is kept independant of any particular scene-graph API. The communication is characterized by command strings across the network. Eg a transformation command string has the form "NSGTRANSFORM [nodeName] [matrix]". Other commands could be added such as file loading, switch, sequences etc. Presently, only the transformation and Cave Navigation command have been implemented.

I have used Limbo to test this API. The diagram below shows a snapshot of the execution. In this example, the Master renders the front wall while the Slaves 1 and 2 render the left and right wall respectively. These are specified using the -wall command-line option defined in CAVElib.


Class Diagrams

Here is the class diagram and description of the classes implemented.

Abstract class that encapsulates a command. Each command has an identifier token string and methods to input/output and execute itself. So, for each command(transform, switch etc ) we define, we need to inherit from this class and override these virtual functions.
Encapsulates a transformation command. Contains the transformation as a 16-float array and the associated node-indentifier. It also overrides the base class functions to read/write/execute itself.
This is the command responsible for sharing CAVE Navigation transform between the nodes. The navigation is stored as a 16-float array. This is just a temporary hack. The distributed CAVElib should be used to share tracking/navigation.
A Singleton object with a map associating a command identifier with its implementation of NsgCmd. Eg the command "NSGTRANSFORM" is mapped to NsgTransformCmd. The registry has functions for lookup/registering commands. The NsgCmd's, on creation, register themselves automatically with the registry.
Inherited from db Shared state. Observes a CAVERNdb_client and is triggered by received network commands. It used the first token of the received string to look up the respective NsgCmd in the registry, and it then makes the command execute. It is also responsible for placing commands on the network.
Abstract class that has functions that execute each NsgCmd. Every NsgCmd has an instance of this class. When the NsgCmd is executed by the NsgCmdInterpreter, it calls the respective functions in the NsgInterface. Eg the NsgTransformCmd would call the doTransform function. Your application should implement this NsgInterface and override all the commands, if not using NsgPfInterface.
While all the classes so far have been independant of any scene-graph API, this is the class that serves as a bridge between a performer app and NSG. Using this class in your app shields you from the rest of the NSG classes. You need to attach the pfNodes that you want to share and call process every frame. Any updates to those nodes will be synchronized

API Docs

Limbo example

This section explains the changes made to limbo in order to use NSG. One instance acts as the master while the rest are slaves(which is by default). To act as master, the line "MASTER" should be added in the .lbo config file.
CAVERN_perfLimbo_c has a new member - master which is true if this application instance acts as the master, false if slave. This is read from the config file.
This general purpose API obviates the need for pfNetDCS. pfDCS should be used instead. You need to create an instance of NsgPfInterface as shown below.
class limboSaber_c : public CAVERN_perfLimbo_c
	bool master;
	NsgPfInterface* nsg;

The nodes that need to be shared need to be registered with nsg. In the case of Limbo, the 3 sabers, the wand dcs and the CAVE Nav dcs are attached. At every frame. the process function for NsgPfInterface is called. Since only the master handles any user-interaction, that should also be checked for in the limboSaber_c::process.

void limboSaber_c::init( int argc, char** argv ) 
	<-- snipped -->
	nsg = new NsgPfInterface(modelClient,"Limbodb","nsg");

void limboSaber_c::process()
	// Give it a few cycles to stop messages from queueing up.
	for(int i = 0; i < 5; i++) modelClient->process();
	if (master) {


  1. Instead of having to add all the nodes to be shared, only the scene should be registered . On process, NsgPfInterface should traverse the scene and depending on the pfNode types use the appropriate NsgCmd's for transmission across the network. The nomenclature we discussed earlier will be used to identify the nodes.
  2. Currently, I use the CAVElib option -wall to specify the wall each instance should render. Should be able to specify no of rendering nodes and arbitrary wall/viewport dimensions for each node in the config file.
  3. Integration with distributed CAVElib to share tracking and navigation.
  4. Address collaboration between applications and latejoiners


Coming soon...

Shalini 6/11/2001