Shiwen Mao

GPS analysis of multiple sessions with applications to admission control

We introduce a statistical method to analyze multiplexing of multiple sessions sharing link bandwidth using generalized processor sharing (GPS) scheduling. Our method is shown to substantially improve previous upper bounds for GPS scheduling of Markov modulated fluid processes (MMFP) sessions, especially as the number of sessions increases. Application of analytical results to admission control indicate that by sharing bandwidth using GPS among traffic classes there are significant gains over systems that statically segregate the link bandwidth. This effect is quantified in several experiments where various combinations of source types are used. The gains are pronounced when bursty sources with stricter QoS requirements are used.