Recommendations for Efficient Implementation of RPL

RPL is a routing protocol that is applicable in a wide range of settings in networks characterized by low power and lossy links (LLN). Because RPL is designed to work in a wide range of settings, it offers many configuration parameters and choices in how different mechanisms are implemented. This flexibility is essential to ensure the wide applicability of this protocol. One can take advantage of this flexibility to implement and configure RPL in the most efficient way for a given network. However, it is easy to inadvertently configure RPL to work inefficiently in the network. These design choices must be made carefully drawing on implementation and operational experiences. In this document, we describe aspects of configuration and mechanisms that impact the performance of RPL. We hope these descriptions serve as guidelines and best practices for RPL implementers and enables them to understand why certain design and configuration choices are favored over others.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119. This terminology used in this document is consistent with the terminologies described in , , and . This document does not introduce new terms.

The minimum Trickle interval determines the fastest rate at which RPL will send DIOs. It is not useful to have multiple DIOs in the transmit queue at a given node. The information in the older DIOs is likely already stale when the new DIO is generated. In systems that cannot cancel the packets that are already in the queue, it is advisable to set the minimum interval to be much larger than the minimum link layer packet time.

The maximum Trickle interval determines the slowest rate at which RPL will send DIOs. It is recommended that the maximum interval is set to several hours. A large interval does not necessarily make RPL less agile or the routing information stale. Trickle will operate at a rate between the minimum and maximum interval depending on the dynamics in the network.

If a node receives more DIOs than the redundancy constant, it does not transmit, i.e., suppresses, its DIO. The rationale for this suppression is that the additional DIOs do not help discover new or better paths if certain number of DIOs have already been transmitted in the neighborhood of a node. In general, the smaller this number the more efficient the route discovery. Setting this value too small can lead to network partitioning as many nodes will suppress their DIOs and will not be discovered. A constant of 3-5 has been found adequate in deployments.

It is often not necessary for a node to poison a route explicitly by advertising a rank of INFINITY. With datapath validation, it is easy to detect a loop and coupled with adaptive beaconing, the routes can be repaired quickly without additional explicit mechanism for route poisoning. Poisoning the route does not prevent loops because the control packet can get dropped on the lossy link.

The neighborhood information is useful even when a node detects that it has lost a route. It is recommended that the nodes not flush the entire or subset of the neighbor table even when a node loses its route or detects a loop. It is sufficient to mark the nodes in the table with the updated information that resulted in route loss or loops, e.g., marking the particular parent with a rank of INFINITY.

When a node detects that a path is inconsistent through datapath validation, it tasks the control plane to repair the topology and make it consistent. During this time, although the route is available, it is advisable that the data packets are sent at lower rates to reduce contention with the control packets. This slow-down can increase data packet latency or lead to queue overflow.

With bursty links, a link metric designed to reflect link quality accurately can change rapidly. Other link metrics may also change rapidly. As a result, the path cost computed using these agile metrics can change rapidly. Selecting the best path then implies frequent parent changes. Route flapping is not detrimental to the performance of many network protocols such as sensor data collection over UDP. Hence, oftentimes, it is better to optimize for path cost than for path stability.

A naive implementation of storing mode will have large state overhead, especially in large networks. However, it is possible for storing mode to use RAM efficiently by state aggregation, compression, and other techniques. Current open source implementations are known to limit the routing table size to 30 (TinyOS on TelosB which has 10KB of RAM) and 100 (Contiki on MSP430F5438-based platform which has 16K of RAM).

Thanks to Ulrich Herberg and Mukul Goyal for valuable comments.

None.

Security considerations to be developed in accordance to the output of the WG.