In modern IP networks, processing cost in network nodes is considered as a bottleneck. This problem is tackled with traffic based IP switching. The performance of traffic based IP switching depends heavily on flow classification. We demonstrate a method to evaluate the performance gains available with this technique with an optimal nondeterministic classifier giving a practical lower bound for processing cost and compare it with two real life classifiers using recorded traces. Keywords IP switching, flow classification, cost optimization, Internet, traffic measurements