Page 42 - Data Science Algorithms in a Week

Using Deep Learning to Configure Parallel Distributed Discrete-Event Simulators 27

calculation. Once all of the nodes have processed events up to their local event horizon,
they are then ready to synchronize. The next step is to compute the global event horizon
as the minimum local event horizon across all nodes. Once GVT is determined, all events
with time stamps less than or equal to GVT are committed (Steinman, Nicol, Wilson, &
Lee, 1995).

Figure 2: The Event Horizon and GVT.

A potential problem is that some of the nodes may have processed events that went
beyond GVT. An event processed by the respective simulation object must be rolled back
when a newly generated events is received in its past. Rollback is very simple in this case
and involves discarding unsent messages that were generated by the event and then
restoring state variables that were modified by the event. Therefore, antimessages are not
required because messages that would create bad events are never released (Steinman
1996).

Breathing Time Warp (BTW)

BTW is another hybrid algorithm for time management and event synchronization
that tries to solve the problems with TW and BTB (Steinman, 1993):

 TW has the potential problem of rollback and cascading antimessage explosions.
 BTW has the potential problem of a higher frequency of synchronizations.