A Measurement based Analysis of the Interaction between Network Layers in TinyOS Umberto Malesci1 2 and Samuel Madden1 1 MIT CSAIL Cambridge MA 02139 umberto madden csail mit edu 2 Fluidmesh Networks Inc Boston MA 02111 umberto malesci fluidmesh com Abstract There have been a number of recent proposals for link and network layer protocols in the sensor networking literature each of which claims to be superior to other approaches However a proposal for a networking protocol at a given layer in the stack is typically evaluated in the context of a single set of carefully selected protocols at other layers Because of the limited data available about interactions between different protocols at various layers of the stack it is difficult for developers of sensor network applications to select from amongst the range of alternative sensor networking protocols This paper evaluates the interaction between several protocols at the MAC and network layers measuring their performance in terms of end to end throughput and loss on a large testbed We identify some common sources of poor performance based on this experience we propose a set of design principles for the designers of future interfaces 1 Introduction The sensor network community has proposed a number of different networking protocols for routing 14 21 media access 4 5 8 16 17 and power management 6 8 16 17 Despite significant innovation in each of these areas there has been little work addressing the interaction of protocols across areas into a single network stack Published papers typically propose changes in one abstraction e g a new MAC layer while using some default implementation of the other abstractions e g using the standard multihop routing protocol in TinyOS and evaluating on a particular topology and a particular workload This approach has led to a large number of competing protocol proposals which are difficult to compare with one another due to varying choices made by authors about appropriate defaults application workloads and network topologies This makes it hard for an application designer to select the best set of protocols for his or her application and impossible for other researchers to understand whether claimed differences between protocols are simply due to artifacts in one experimenter s setup or are true differences between protocols In this paper we focus on the TinyOS 7 operating system because source code for many different protocol implementations is widely available and because it appears to be the current platform of choice for sensor network research In TinyOS problems related to the interactions between protocols are aggravated by significant disagreement in the community about how functionality should be spread across different network layers Lines between layers are blurred making innovation difficult and mixing and matching of implementations tricky as interfaces are poorly specified As an example consider per link acknowledgments and retransmissions which are widely regarded as an important feature for reducing loss in sensor networks Different designers implement these features in different parts of the network stack Retransmissions due to negative or failed acknowledgments can be implemented at either the link or network layers as can duplicate suppression Placement of these operations is not consistent across implementations for example the B MAC 17 link MAC protocol implements acknowledgments but assume that retransmissions and not duplicate suppression should be implemented in the layers above whereas the S MAC protocol implements all three If a researcher wants to design a new network layer he must choose which of ACKs retransmissions and duplicate suppression to implement These choices will invariably tie his implementation to a particular MAC layer and limit the generality and impact of his work worse yet due to unstated assumptions in various implementations he may believe his protocol will work with a particular MAC layer only to find it does not If the abstraction boundaries between layers were more cleanly implemented and specified these issues would not arise In this paper we present a systematic study of the performance in terms of network loss rate of different combinations of MAC routing forwarding and power management protocols that have been previously proposed in the literature Our aim is to provide measurements that will enable a first step towards fixing these problems with the network protocols in TinyOS by 1 Benchmarking several widely published protocols with the same application workload and on the same network topology 2 Highlighting the significant differences between different implementations protocols that are ostensibly at the same layer 3 Illustrating a number of examples where interactions between different layers lead to significant performance degradation 4 Recommending combinations of protocols for particular application workloads 5 Illustrating that no one protocol at any of these layers strictly dominates any other protocol despite claims to the contrary in the literature and that there are many hidden assumptions and subtle issues with even widely used network protocols The purpose of this paper is not to proscribe a specific layering or to suggest that one implementation is better than another Rather we aim illustrate some of the limitations of the current state of software in TinyOS so that the community can move towards a cleaner set of interfaces that support greater protocol diversity and allow application developers to make more informed choices about the appropriate selection of networking protocols 2 Sensor Network Architecture A common architecture and set of abstractions has emerged for the network stack for wireless sensor networks As in most network architectures the basic abstraction is the layer However the TinyOS network stack differs from the traditional Internet stack in several ways Layers make abundant use of cross layering in order to increase throughput and decrease power consumption 10 Power management is present in many different forms in several layers The network stack in TinyOS can be broken into four major layers the physical layer the link MAC layer to keep consistent with the naming using in many publications we refer to this simply as the MAC layer in the remainder of this paper the forwarding routing layer we refer to this as the routing layer in the remainder of this text and the application layer Moreover the network stack