Compute Server

ComputeServer is an AssociativeMemory idiom (or pattern composite) that allows multiple distributed Masters to submit task and gather results computed by multiple, distributed, and possibly parallel workers. A Master submit tasks to the AssociativeMemory. Workers blindly grab Task objects from the AssociativeMemory, send them the #run method and store its answer back into the AssociativeMemory. The Master can then retrieve this result. The Task object is usually some implementation of the CommandPattern. This is a simplification of the BlackboardMetaphor.

Sample:

This is an example of a simple ComputeServer using JavaSpaces. First, we will define the basic TaskEntry. JavaSpaces requires the use of concrete classes that implement the net.jini.core.entry.Entry interface. Since we want to use it as an abstract class, we will define its run method to throw a subclass responsibility exception:

 import net.jini.core.entry.Entry;

 public
     class      TaskEntry
     implements Entry
 {
     public ResultEntry run()
     {
         throw new RuntimeException( "Subclass Responsibility" );
     }
 }

The ResultEntry is even simpler:

 import net.jini.core.entry.Entry;

 public
     class      ResultEntry
     implements Entry
 {
 }

The worker is pretty straightforward -- it looks for TaskEntry objects , removes them from AssociativeMemory, sends them the #run message, and write the ResultEntry back to AssociativeMemory. Here's some basic code:

 import net.jini.core.entry.Entry; 
 import net.jini.core.lease.Lease;
 import net.jini.space.JavaSpace;

 public
     class      Worker
     implements Runnable
 {
     public void run()
     {
         JavaSpace space = SpaceAccessor.getDefault();
         TaskEntry pattern = new TaskEntry();

         for (;;)
         {
             try 
             {
                 TaskEntry aTask = (TaskEntry) 
                    space.take( pattern, null, Long.MAX_VALUE );

                Entry result = aTask.run();
                if ( result != null )
                    space.write( result, null, Long.MAX_VALUE );
             }
             catch ( Exception e ) 
             {
                 log( e );
             }
         }
     }
 }

This can be started from a thread, an exec-process, or from a main. An actual implementation would include leased entries and transactions. We can also create an abstract Master that generates tasks and collects results on discrete threads:

 public abstract class Master
 {
     private JavaSpace m_space;
     private Thread m_generator;
     private Thread m_collector;

     public Master( JavaSpace space )
     {
         m_space = space;

         m_generator = new Thread() {
             public void run() {
                 generate(); } };

         m_collector = new Thread() {
             public void run() {
                 collect(); } };
     }

     public void start()
     {
         m_generator.start();
         m_collector.start();
     }

     public JavaSpace getSpace()
     {
         return m_space;
     }

     abstract public void generate();
     abstract public void collect();
 }

Now that we have the basic infrastructure in place, we can start using our compute server by defining concrete Task, Result, and Master types.

 public class MandelTask extends TaskEntry
 {
     public Complex x, y;
     ...

     public RsultEntry run()
     {
         calculate next in mandelbrot in set...
     }
 }

 public
     class   MandelResult
     extends ResultEntry
 {
     ...

     public void draw()
     {
     }
 }

The more worker processes we add to the ComputeServer, the faster the set will be calculated. Of course, this is just an example and the distributed overhead may outweigh the benifits of the compute server. The master for this would look something like the following:

 public
     class   MandelMaster
     extends Master
 {
     public MandelMaster()
     {
         super( Space.getDefault() );
     }

     public void generate()
     {
        while ( complex points )
        {
            getSpace().write( new MandelTask( params 0 ) );
        }
     }

     public void collect()
     {
        Entry pattern = new MandelResult();

        while ( more to collect )
        {
            MandelResult result = (MandelResult)
                getSpace().take( pattern, null, Long.MAX_VALUE );

            result.draw();
        }
     }
 } 

When you call master #start, the collector thread will start plotting the results on the screen as soon as they are generated.

---

A good example of the ComputeServer pattern is the SetiAtHome screen saver project (http://setiathome.ssl.berkeley.edu/sciencepaper.html). SetiAtHome uses multiple, autonomously collaborating unskilled workers (your computer in screen saver mode) that continually grab tasks (figures to be analyzed), process them, and return the results to an AssociativeMemory (i.e. SETI).

Processes posting tasks, processes retrieving results, and processes that are removing and executing Tasks from AssociativeMemory operate collaboratively but autonomously -- they aren't aware of each other. One can bring on new workers without disturbing the system. Adding a new processor will usually result in increased performance. You also see this style used a lot in crypto-analysis. Unlike the BlackboardMetaphor, the algorithm is specialized within the actual Task Tuple. -- RobertDiFalco (this comment was edited and moved from TupleSpace)

---

See also: GenerativeCommunication