Analysis of system performance for differentiated handling of SIP requests

Abstract

The IP multimedia subsystem (IMS) relies on the session initiation protocol (SIP) protocol for call management and signaling. The SIP protocol defines several types of request messages for its operation. A call setup, for instance, involves SIP INVITE request, which often generates message exchanges across multiple SIP proxies before the call is established. While requests such as INVITE are initiated by end users (referred herein as user-triggered requests), other SIP requests are system-triggered, wherein the user equipment (UE) automatically generates certain SIP requests for addressing and other housekeeping purposes. In particular, the SIP REGISTER requests are required to allow the UEs to register themselves with designated SIP registrar servers. Also, UEs are required to refresh their registrations by means of fresh REGISTER requests, at regular intervals. Consequently, a SIP server can experience a significant amount of load, even if the user call rates are low. In this paper, we suggest that the performance of SIP based systems can be improved by differentiated handling of the user-triggered and system-triggered requests. For instance, since the user-triggered requests have more direct impact on response time experienced by end users, these type of requests can scheduled with a higher priorities on the server. Moreover, the system-triggered requests can be delayed or even curtailed, in order to ldquosmoothenrdquo the request arrival rates at the server. We also attempt to quantify the impact of user-triggered and system-triggered requests, on the performance of an IMS system. In particular, we compute the performance of IMS by means of simulations, driven by real-world traffic statistics. We further propose and analyze a simple heuristic algorithm for controlling the rate of generation of REGISTER requests. In our simulation experiments, we find that the average server load, as measured by its job queue size, decreased by 50% when the algorithm was applied.