An Experimental Investigation of Issues in End-to-End QoS

Meaningful Quality of Service (QoS) guarantees for`delay sensitive' networked applications must be end-to-end. This paper presents an experimental study of applications using computer workstation end-points. For the application experiments, the variables under study were the degree of multiprocessing, the transport protocol used (UDP or TCP), the process structure used by the application, and the eeectiveness of priorities for achieving \real-time" behavior. The experiments were performed on the IBM RISC System/6000 workstation running AIX 3.2.5. While worst-case network delays can dominate any factors at work in the end-points, our experiments show that in the common case of a lightly loaded LAN, operating system eeects are of greater magnitude than any jitter introduced by the network. Operating system issues are thus shown to be as challenging as network innuences for end-to-end QoS provision. For our experimental variables, we found that: (1) priorities are of limited usefulness, (2) application process structure strongly aaects jitter introduced by the scheduler, (3) the transport protocols produce diierent delay and jitter tradeoos, and (4) the degree of multiprocessing has a large eeect both on delay and jitter. The experimental results suggest that improved real-time support in workstation operating systems is essential for end-to-end QoS provision. Further, it is clear that this support requires careful coordination among the application, network and operating system.