A Framework for Evaluating Server Performance: Application to SIP Proxy Servers.

KRISHNAMURTHY, RAMESH . A Framework for Evaluating Server Performance: Application to SIP Proxy Servers. (Under the direction of Dr. George N. Rouskas.) The growing number of applications that use the Session Initiation Protocol (SIP) to manage media sessions over IP is placing increasing demands on the SIP proxy servers (SPS) that make up the core of the SIP network. In this work we investigated the performance of OpenSIPS, an open source SIP proxy server. We collected a large set of experimental data to characterize the performance of the SPS under various call arrival rates and inter-arrival time distributions. Based on these measurements, in the first part, we studied a single SPS server thread on a single-core CPU hardware and modeled the SIP proxy server as an M/G/1 queue. A key component of the model is a parameter that captures the interrupt overhead, i.e., the impact of interrupts and the resulting cache-misses on socket queue service times. For the second part, we studied the performance of multiple SPS server threads on a singlecore CPU hardware. We measured the call rate where the SPS server starts experiencing losses greater than 1% and developed a prediction model for the drop probability as a function of call rate and number of server threads. We also introduced a new parameter to capture the overhead of multiple server threads, in addition to the interrupt overhead. For the third part, we investigated the impact of the Linux scheduler settings on the performance of single-core, multi-threaded SIP proxy servers, in terms of packet service time, waiting time, and packet drop rate (PDR) to capture the impact on end-user experience. Based on the results of a large set of experiments across a wide range of traffic loads and number of server threads, we developed a methodology to configure the scheduler parameters such that it resulted in significant gains in SPS performance compared to industry-recommended “server” mode operation. Importantly, the gains in performance were the result of setting the scheduler parameters to appropriate values, without the need for adding server capacity or other capital expenditures. For the final phase of our research, we investigated the impact of the Linux scheduler’s load-balancing algorithm on the performance of multi-threaded SIP proxy server running on a multi-core processor system. We conducted extensive experiments and developed a practical guidelines for tuning the scheduler to a “enhanced multi-core server mode” that results in significant gains in performance, thus addressing one of the most crucial needs in today’s data center: to extract performance gains from the existing computing infrastructures. We further developed a capacity planning model that provides a good first-order approximation of the total capacity of the SPS system in terms of the call arrival rate that may be supported without affecting user experience in terms of dropped call. Our measurement and modeling methodology is general, and can be applied to characterize the performance of a wide range of network application protocols.