What's new in WARPED

Three types of dynamic memory managers have been added to the simulation kernel. In addition, five major features have been added to the simulation kernel code.

• Feature One (Removal of containers from input and state queue): The inputQueue is now a queue of BasicEvents and the StateQueue is a queue of BasicStates. The output queue still retains its original structure and is a queue of containers with BasicEvent pointers.

  
  

    

  Figure: Pointers in the state

  
  
  
• Feature Two(Introduction of queue pointers in the state): The second optimization was introduced to reduce the searching in the queues during rollback (this was found to be a major time consuming function in the WARPED kernel). The basic idea is to have two pointers in the state. One pointer to point to the current position in the input Queue, and the other to point to the tail of the output Queue. Upon rollback, the correct state is restored and from the restored state, we can determine where to rollback in the inputQueue and in the output queue. Figure  illustrates the new structure of the queues. This significantly reduces the search time required in the queues during rollback. In addition, garbage collection also works off the queue pointers in the state. When GVT crosses the time of a state, garbage collection routines are invoked on that state. This in turn will invoke garbage collection routines on the input and the output queue. As the state contains pointers to the input and output queue, garbage collection of the input and output queues becomes trivial.
• Feature Three (One antimessage per rollback): Only one antimessage is sent out upon rollback. The lowest timestamp antimessage is sent out and the other antimessages are suppressed. At the receipient's end, upon receiving an antimessage, all messages with timestamps greater than that of the antimessages are deleted. This greatly reduces the total number of messages that are generated in the system and improves performance.

  
  

    

  Figure: Message Aggregation in WARPED

  
  
  
• Feature Four(Message Aggregation capability): A MessageManager module has been built on top of the MPI communication library. This functionality has been added to the CommManager module. Figure  illustrates the basic scheme. In message passing environments, the message send time is dominated by overheads that are relatively independent of the message size. Therefore, fine-grained applications (such as TimeWarp simulators) suffer high overheads because of frequent communication. This optimization optimizes the communication subsystem of Time-Warp simulators using dynamic message aggregation . Under this scheme, Time-Warp messages with the same destination LP, occuring in close temporal proximity are dynamically aggregated and sent as a single physical message. Several aggregation strategies that attempt to minimize the communication overhead without harming the progress of the simulation (because of messages being delayed) have been developed.
• Feature Five (Terminate Simulation capability): A TimeWarp simulation object can now terminate simulation by calling the terminateSimulation(char *) method. When a TimeWarp object calls the terminateSimulation(char *) method, the logical process (LP) on which the object resides, sends a TERMINATE message to all other LPs. But be forewarned that even though a call to terminateSimulation(char *) causes the simulation to terminate cleanly, it can violate causality in certain special cases and potentially cause an incorrect simulation. The use of this function is left to the user's discretion and it is the user's responsibility to make sure that an incorrect simulation does not occur.