Flexible. Efcient and Robust Real-Time Communication with Server-based Ethernet Switching * R. Santos, A. Vieira, P. Pedreiras, A. Oliveira DETI/ IEETA / University of Aveiro Aveiro, Portugal { rsantos,alexandrevieira,pbr,amaldo } @ua.pt Thomas Nolte L. Almeida, R. Marau IEETA - DEEC / University of Porto 4200-465 Porto, Portugal { lda,marau} @fe.up.pt MRTC / Malardalen University Vasteras, Sweden thomas.nolte@mdh.se Abstract The inormation exchanged in Networked Embedded Systems is steadily increasing in quantit, size, complex it and hetergeneit, with grwing requirements for ar bitrr arrival patters and guarnteed QoS. One of the networking prtocols that is becoming more common in such systems is Etheret and its real-time Etheret vari ants. However the hardly support all the referred require ments in an efcient manner since the either favour deter minism or thrughput, but not both. A potential solution recently proposed by the authors is the Server-SE prto col that uses servers to confine trfc associated to specic applications or subsystems. Such an apprach is dynam ically reconfgurble and adaptive, being more bandwidth efcient while prviding composabilit in the time domain. This paper prposes integrting the servers inside the Eth eret switch, boosting both thefexibilit and the rbustness of Server-SE, allowing, for example, the seamless connec tion of any Etheret node. The switch is an FI-enabled Etheret Switch and the paper discusses two specic ways of integrting the servers, namely in soware or in hard ware. These options are described and compared analyti cally and experimentally. The former favours fexibilit in the servers design and management while the latter pr vides lower latency. 1. Introduction There has been a continued steep increase in the com plexity, quantity and heterogeneity of the data exchanged between nodes in Networked Embedded Systems (NES). From data originated in simple 10 bit ADCs to multi kilobyte variable bit-rate multimedia traffc. Moreover, many NES are frequently subject to real-time constraints 'This work was partially supported by the iLAND project, call 2008- 1 of the EV ARTEMIS JV Programme, by the European Community through the ICT NoE 214373 ArtistDesign and by the Portuguese Gover ment through the FCT project HaRTES - PTDCIEEA-ACRJ7330712006 and Ph.D. grant - SFRHlD/3281412006. 978-1-4244-5461-7/10/$26.00 ©2010 IEEE that extend to the respective information exchanges, requir ing support from a real-time network. One network tech nology that became widely used in these systems is Eth ernet [1], which conquered the offce automation market long ago, entered massively into the factory automation and large embedded systems domains and is now being consid ered for mass market domains such as the automotive one. However, Ethernet was not originally developed to meet the requirements of NES, namely in what concerns key aspects such as predictability and timeliness. These limitations led to the development of the so-called Real-Time Ether net (RTE) protocols, but even these still reveal difculties in handling the variety of requirements that current NES pose in an efcient manner, particularly arbitary arrival patterns and widely diferent QoS requirements. Typically, such protocols were either tuned to achieve high bandwidth efciency or strict timeliness guarantees but not both. Standard Ethernet switches are typically designed for high throughput Internet access or fle sharing, presenting limitations in what concerns real-time performance, with potentially long queueing delays or even packet losses re sulting from limited scheduling capabilities (up to 8 prior itized FIFO queues, only) and a generalized lack of mem ory partitioning. The techniques proposed to overcome such limitations range from shaping the tafc submitted to the switch [2] to limiting that trafc by application de sign [3], adding transmission control features [4] and pro viding more efcient scheduling policies and admission control [5] [6] [7]. Amongst the most proeminent market contenders we can fnd EtherCAT [4], PROFINET [6], Eth ernetIP [3] and TTEthernet [7]. One of the main challenges in designing current NES is managing the ever growing level of complexity [8] [9] [10]. Component-oriented design methodologies, which pro vide composability, are paticulaly well suited since they support safe resource sharing, allowing different compo nents/subsystems to be developed separately and later in tegrated in the system. On the other hand, server-based scheduling is recognized as an efective means to enable such kind of resource sharing [11] and it can be the ba sis for resource partitioning and virtualization, supporting