Stochastic Systems Group  

Tracking Transaction Using Timestamps
Anima Anandkumar
SSG, MIT
The issue of tracking using timestamps arises in many enterprise systems which lack proper monitoring instrumentation and only have logrecords with timestamps.
We consider tracking transactions in a queuing network using timestamps generated at the network nodes. In general, the optimal maximum likelihood tracking has exponential complexity in the number of transactions. However, we prove that maximum likelihood tracking reduces to a series of bipartite minimum weight perfect matchings for the case of acyclic infiniteserver queuing networks. Hence, maximum likelihood tracking is equivalent to decentralized matching of timestamps in small clusters of network nodes. Further, we show that the tracking system has least accuracy under exponential or uniform services for a given set of timestamps.
Problems with this site should be emailed to jonesb@mit.edu