| 9 | | The past decade has seen significant progress towards the goal of making long-term, embedded, environmental sensing a reality. Within a typical mote-class device, current technology dictates that the vast majority of energy is consumed by the physical radio. As such, duty-cycling of the radio has been the key method employed to reduce energy consumption in practical deployments.~\cite{gdi04sensys,tolle05sensys,firewxnet06mobisys,sensorscope08ipsn} A number of ultra-low power MAC layers have sustained the race to zero consumption at the link layer\cite{ye06sensys,dozer07ipsn}, where the current state-of-the-art has a duty cycle of 0.65\%~\cite{hui08sensys}). However, we believe that with current technology, there is minimal further improvement possible for duty-cycling link layers. |
| | 9 | The past decade has seen significant progress towards the goal of making long-term, embedded, environmental sensing a reality. Within a typical mote-class device, current technology dictates that the vast majority of energy is consumed by the physical radio. As such, duty-cycling of the radio has been the key method employed to reduce energy consumption in practical deployments.~\cite{gdi04sensys,tolle05sensys,firewxnet06mobisys,sensorscope08ipsn} A number of ultra-low power MAC layers have sustained the race to zero consumption at the link layer\cite{ye06sensys,dozer07ipsn}, where the current state-of-the-art has a duty cycle of 0.65\%~\cite{hui08}). However, we believe that with current technology, there is minimal further improvement possible for duty-cycling link layers. |
| 11 | | A common thread among most of the deployments mentioned is an implicit requirement that the network is 'always available' - the link layer provides the ability to interact with any node in the network at any time. Though this provides a degree of comfort for network operators, it often limits the capacity of the network to make sensor measurements by monopolizing energy resources. |
| 12 | | |
| 13 | | To get a sense for how much the radio dominates an example environmental sensing application energy budget, we show two pie charts in Figure~\ref{fig:energy}: (a) shows the energy distribution between the hardware components for a node running the Low Power Listening (LPL) radio duty-cycling MAC layer~\cite{polastre04} and (b) shows the same distribution for a node that runs the same LPL layer, but only 10\% of the time; the other 90\% of the time is spent with the radio completely off. Though the radio in each case is the majority consumer, the magnitude of consumption is nearly an order of magnitude less in the latter. This provides an opportunity to reassign the Joules previously reserved for radio idle listening to more useful tasks like increased sensing. |
| 14 | | |
| 15 | | Two challenges arise from this approach - the need to reconstruct routing links and trees after waking up and the increased cost of transmission due to batching. First, examining network reconstruction |
| 16 | | |
| 17 | | Even after radio duty-cycling, radio is still one of the major energy consumers of sensor nodes. As shown in Figure~\ref{fig:energy}(a), for a node running LPL (512ms sleep interval) and transmitting every minute, radio consumes over 95\% over the overall average power draw of 620$\mu$A. In other words, a significant energy cost is paid to allow the node to be in a state of ``always on'', and only a small part of the energy is used in the actual sampling and transmission of information\cite{prabal07batch}. |
| | 11 | A common thread among most of the deployments mentioned is an implicit requirement that the network is ``always available'' - the link layer provides the ability to interact with any node in the network at any time. Though this provides a degree of comfort for network operators, an examination of an environmental sensing application energy budget reveals that network responsiveness has implicitly been allocated as the highest priority for the network. To get a sense for this, we show two pie charts in Figure~\ref{fig:energy}: (a) shows the energy distribution between the hardware components for a node running the Low Power Listening (LPL) radio duty-cycling MAC layer~\cite{polastre04} and (b) shows the same distribution for a node that runs the same LPL layer, but only 10\% of the time; the other 90\% of the time is spent with the radio completely off. Though the radio in each case is the majority consumer, the magnitude of consumption is an order of magnitude less in the latter. This provides an opportunity to reassign the joules previously reserved for radio idle listening to more useful tasks like increased sensing and collection. |
| 31 | | An obvious way to reduce the amount of energy consumed by idle listening is to turn the radio off. Whilst the allows a large amount of energy to be redistributed to tasks such as sampling and sending (when the radio is switched on again), this approach allows incurs an additional network overhead each time the radios are turned back on in where network routing tables must be reformed. Figure~\ref{fig:energy}(b) illustrates the |
| 32 | | clear reduction in energy that comes about when the radio is off for long-periods (in this case 90\%), where the total consumption is reduced, in this example, to 87$\mu$A. |
| 33 | | |
| 34 | | %nature of this additional cost, showing the relationship between time the radios are off and the effective energy consumed per bit of data transmitted. Once radios are off long enough, the effect of amortizing the cost of updating the network state over long periods becomes clear where in these cases the net energy cost is less than a typical low-power listening (LPL) MAC \cite{polastre04}. |
| 35 | | |
| 36 | | The key trade-off in turning off radios is that of network responsiveness. Once nodes only become active in scheduled intervals then the times a user can communicate with nodes, or the times which nodes can report to a base are now constrained. This has clear implications for event-driven applications or query-based systems, where a user may want an immediate responsive from the network. |
| 37 | | |
| 38 | | \subsection{Motivation}~\label{sec:motivation} |
| 39 | | The balance between network responsiveness and the energy consumption dedicated to actual sensing/sending is an important one that has been under-explored in the community to date. The ``always on'' philosophy has dominated sensor networks systems such that (when taking into account the relative energy consumption of listening compared to sampling), network responsiveness has implicitly been allocated as the highest priority for networks. |
| 40 | | |
| 41 | | We argue that for the vast majority of environmental sensing applications, network responsiveness should not be the highest priority role for networks. In order to reduce this implicit bias towards network responsiveness, we argue that moving towards a ``radio off'' model is far more appropriate for most applications. By redistributing energy consumption away from idle listening, we can more freely control the relationship between energy consumption in a more flexible network \emph{utility}. Furthermore, we believe the ``slow-and-steady'' approach to environmental sensing, where nodes consume as little energy as possible, is the wrong paradigm. We argue that consuming as \emph{much} energy as possible, within appropriate energy constraints, allows for networks to perform a far more useful function over the life of the network. |
| | 25 | The balance between network responsiveness and the energy consumption dedicated to sensing and collecting data is an important one that has been under-explored in the community to date. In this paper, we relax the assumption that the network is always available and examine some of the new tradeoffs that are revealed. By redistributing energy consumption away from idle listening, we can more freely control the relationship between energy consumption in a more flexible network \emph{utility}. Furthermore, we believe the ``slow-and-steady'' approach to environmental sensing, where nodes consume as little energy as possible, is the wrong paradigm. We argue that consuming as \emph{much} energy as possible, within appropriate energy constraints, allows for networks to perform a far more useful function over the life of the network. |
| | 34 | We begin with a brief comment on how the paradigm of relaxed responsiveness affects energy consumption by the radio. As has already been discussed, turning the radio off results in a significant reduction in energy consumption. However, after the radio is turned back on, there is some \emph{increased} consumption due to the extra traffic needed to reconstruct a routing tree after waking up and send the sensed data collected while the radio was off. |
| | 35 | |
| | 36 | First, examining network reconstruction, acquiring data on eligible neighbors generally takes less than 5 packet exchanges with each neighbor. In the environmental sensing domain, networks are typically not dense, so neighbor counts are low. Additionally, the design of LPL makes receiving packets relatively cheap, while sending packets is relatively costly. Assuming a large LPL sleep interval of one second, the result is less than five extra seconds of radio operation per link per wakeup, though since wireless is a shared medium, the number of transmissions will likely be less. Since LPL channel checks last 12 ms, this extra wakeup cost is at worst-case roughly equivalent to ~417 LPL channel checks (5000 / 12 ~= 416.66). With our sample LPL interval of one second, that means that as long as we sleep longer than 417 seconds (~7 minutes) per link, we are going to save energy by going to sleep. This improves as the LPL sleep interval is reduced. |
| | 37 | |
| | 38 | Second, concerning sending the extra data collected while asleep, LPL sends a train of packets to ensure that the next hop is guaranteed to receive the message. However, in the non-broadcast case, when any packet is acknowledged, the node ceases to continue sending the packet. Thus, the cost of batching is significantly reduced, as the second and later packets are simply sent immediately after the first while the attention of the next hop has already been claimed. This ability to amortize the costs of acquiring the channel allows batching to actually be more efficient than sending individual packets.~cite{prabal07batch} |
| | 39 | |
| | 40 | The key trade-off in turning off radios is that of network responsiveness. Once nodes only become active in scheduled intervals then the times a user can communicate with nodes, or the times which nodes can report to a base are now constrained. This has clear implications for event-driven applications or query-based systems, where a user may want an immediate responsive from the network. |
| | 41 | |
| | 42 | %An obvious way to reduce the amount of energy consumed by idle listening is to turn the radio off. Whilst the allows a large amount of energy to be redistributed to tasks such as sampling and sending (when the radio is switched on again), this approach allows incurs an additional network overhead each time the radios are turned back on in where network routing tables must be reformed. Figure~\ref{fig:energy}(b) illustrates the |
| | 43 | %clear reduction in energy that comes about when the radio is off for long-periods (in this case 90\%), where the total consumption is reduced, in this example, to 87$\mu$A. |
| | 44 | |
| | 45 | %nature of this additional cost, showing the relationship between time the radios are off and the effective energy consumed per bit of data transmitted. Once radios are off long enough, the effect of amortizing the cost of updating the network state over long periods becomes clear where in these cases the net energy cost is less than a typical low-power listening (LPL) MAC \cite{lpl04sensys}. |
