next up previous
Next: Tolerance and Accumulated Simulation Up: Network Management of Predictive Previous: Introduction to the Predictive

Characteristics of the Predictive Network Management System

There are two types of false messages generated in this predictive network management system; those produced by messages arriving in the past Local Virtual Time (LVT) of an LP and those produced because the LP is generating results which do not match reality. If rollbacks occur for both reasons the question arises as to whether the system will be stable. A stable predictive network management system is one in which rollbacks do not have a significant impact on the system performance. A stable system is able to make reasonably accurate predictions far enough into the future to be useful. An unstable system will have its performance degraded by rollbacks to the point where it is not able to predict ahead of real-time. Initial results shown later indicate that predictive network management systems can be stable.

There are several parameters in this predictive network management system which must be determined. The first is how often the predictive network management system should check the LP to verify that past results match reality. There are two conditions which cause LPs in the system to have states which differ from the system being managed and to produce inaccurate predictions. The first is that the predictive model which comprises an LP is most likely a simplification of the actual managed entity and thus cannot model the entity with perfect fidelity. The second reason is that events outside the scope of the model may occur which lead to inaccurate results. However, a benefit of this system is that it will self-adjust for both of these conditions.

The optimum choice of verification query time, tex2html_wrap_inline775 , is important because querying entities is something the predictive management system should minimize while still guaranteeing that the accuracy is maintained within some predefined tolerance, tex2html_wrap_inline689 . For example, the network management station may predict user location as explained later. If the physical layer attempts spatial reuse via antenna beamforming techniques as in the RDRN project, then there is an acceptable amount of error in the steering angle for the beam and thus the node location is allowed a tolerance. The tolerances could be set for each state variable or message value sent from a LP. State verification can be done in one of at least two ways. The LP state can be compared with previously saved states as real time catches up to the saved state times or output message values can be compared with previously saved output messages in the send queue. In the prototype implemented for this predictive network management system state verification is done based on states saved in the state queue. This implies that all LP states must be saved from the LP LVT back to the current time.

The amount of time into the future which the emulation will attempt to venture is another parameter which must be determined. This lookahead sliding window width, tex2html_wrap_inline693 , should be preconfigured based on the accuracy required; the farther ahead this predictive network management system attempts to go past real time, the more risk that is assumed.




next up previous
Next: Tolerance and Accumulated Simulation Up: Network Management of Predictive Previous: Introduction to the Predictive

Steve Bush
Thu Feb 27 15:34:42 CST 1997