A framework for improving the performance of application servers in next generation networks

Next generation networks (NGNs) such as IP Multimedia Subsystem (IMS) are completely built on the Internet Protocol (IP) suite. This has made IP the de facto standard for data networking, voice over IP (VoIP), and media rich applications such as streaming multimedia, ringtones, multi-player gaming, and high-definition video conferencing for remote interaction. A primary feature of such converged networks is that they use the same IP-based network for simultaneously delivering voice, video, and data. Such services are provided on application servers built using industry standard Advanced Telecom Computing Architecture (ATCA) based blade computing units with various flavors of commodity open source operating systems like Linux, xBSD, and OpenSolaris. However, real-time and latency sensitive applications such as streaming multimedia require that the entire network path of packet delivery from the originating server to the end host be properly and appropriately configured so as to avoid unnecessary delay and jitter in the data transfer mechanisms. With the ease of deployment comes the challenge of delivering such rich multimedia applications in NGNs since there exists no separate paths for voice and data as present in existing circuit-switched public switched telephone network (PSTN). Packet delivery in such converged architectures involves interaction between the storage disks, operating system (OS), network interface cards (NICs), and the various switches and routers—each of which is independently capable of introducting delay in the data transfer mechanism. In this dissertation, we focus on understanding and improving the performance of application servers present in high traffic content delivery networks (CDNs) and hoisting latency sensitive applications with heavy I/O requirements. We start by identifying an architectural framework for traffic characterization that is expected to provide insights about the composition and dynamics (e.g., average packet size and data rate, protocol composition) of network traffic present in CDNs. Once the nature and type of network traffic arriving at the NICs have been identified, we attempt to identify packet processing bottlenecks due to the interaction between the NICs, OS, and the underlying hardware. We propose a closed form queuing model that aims to understand the packet processing capabilities of the NICs based on the available computing resources. We have shown that there exist limits beyond which a computing unit cannot process packets without overloading the CPU. Since the performance of latency sensitive processes can be negatively impacted by delays of the storage network and by the dynamics of the OS, we present solutions for prioritizing the reader processes and tweaking the pagedaemon in open source OS. Based on our implementation in the NetBSD kernel, we have observed an approximate 15%–20% improvement in the transactions per second (TPS) of latency sensitive applications. Finally, we believe that our framework and approach of identifying the basic components in network data transfer mechanisms are for most part generic and can be used for performance tuning and deploying application servers in NGNs with a variety of different services.