On the isolation of several work-conserving scheduling policies

In this paper we study the isolation of five work-conserving scheduling policies in connection-oriented packet-switched networks. We say that a policy has good isolation if its performance (end-to-end packet delay here) is not influenced by the session configuration. Here we study, by simulation on a very simple setup, how the average packet delay changes in one session when the length or number of the rest of sessions change (while the total rate at each link is preserved). In our study we consider two well-known scheduling policies, namely weighted fair queueing (WFQ) and FIFO, a recently proposed label-based policy S-CEDF, and two more label-based policies we introduce here for connection-oriented networks. We observe that the performance of WFQ and FIFO tends to significantly vary under changing environments, while the label-based policies tend to be more stable. In particular, we observe that the end-to-end delay of WFQ and FIFO decreases when the length of the sessions competing with one given increases. When the environment changes by increasing the number of sessions, the delay under WFQ tends to decrease with the divisions, while the delay under FIFO tends to increase. Further work is needed to analyze these results, and to obtain empirical and analytical isolation bounds for a variety of policies