On per-session end-to-end delay distributions and the call admission problem for real-time applications with QOS requirements