Measuring and reducing the impact of the operating system kernel on end‐to‐end latencies in synchronous packet switched networks

Abstract

This paper presents an evaluation of the impact of the so-called operating system (OS) latencies on the performance of a synchronous network based on global time coordination. The concept of end-to-end latency was first defined by extending the concept of latency used to evaluate the performance of real-time systems and the end-to-end latency provided by a general-purpose OS was measured as a benchmark. Finally, real-time techniques were used to reduce the worst-case values of such a latency, showing how a gateway between synchronous and asynchronous networks can be implemented by using commercial-off-the-shelf hardware and a proper software stack (based on a real-time version of Linux). The use of a real-time OS is still a nontrivial task, which requires experience and the analysis of the specific application to devise the proper techniques to be applied. This work dissects the problem of OS-to-network data transfer (and vice versa) identifying the key sources of latencies and delay jitter, and solving each problem with the application of a proper technique.is only possible if packets are sent and received according to an appropriate schedule, and such a schedule is strictly respected. For example, if a router receives at the same time two packets sent to the same interface, it is not able to properly forward them without buffering. Such a strict synchronisation between routers and end systems can be achieved by means of global time references such as Universal Time Coordinated (UTC), or Global Positioning System Time (GPST), the time used by the global navigation system, which is available also for civilian use for free and without restrictions in normal times. Future navigation systems like the European Galileo and the Chinese BeiDou will only improve the availability of a global time reference. Alternatively, several solutions for distributing the synchronisation reference directly through the network can be adopted; see for instance [16,17], and, specifically for TDP, [18].Although the availability of global time empowers global coordination, it does not mean that synchronisation needed for PF comes for free, as it requires the ability, for both end systems and routers, to send packets 'at the correct time' to properly respect the schedule. Specially considering the end systems, this may be a challenge because of the so called 'latency', which can be informally defined as the difference between the time when a packet should be received (sent), and the time when it is actually received (sent). Whereas modern hardware (even off-the-shelf hardware) is generally able to provide the low latencies needed for PF, the software stack is more problematic, as the latencies generated by an operating system (OS) kernel can be large enough to compromise the correct forwarding of packets or to force to use a low bitrate, as it will be shown in Section 3.Whereas some works related to high-performance network applications only consider average throughput and do not care about the latencies introduced by the OS [19,20], it ha...

EMP: execution time measurement protocol for compute‐bound programs

“…A variety of works take an empirical approach to different aspects of software. Some perform an empirical study related to measuring software performance . Others empirically investigate the performance of an algorithm written in C or in Java .…”