CROSS REFERENCE TO RELATED APPLICATIONS

This application is a continuation in part of U.S. application Ser. No. 10/922,371, titled “Method for Operating a Telecom System,” filed Aug. 20, 2004, which is hereby incorporated by reference, in its entirety. This application also claims priority to U.S. Provisional Application No. 60/617,897, titled “Method for Operating a Combined MultimediaTelecom System,” filed Oct. 12, 2004, which is hereby incorporated by reference, in its entirety.
BACKGROUND OF THE INVENTION

1. Field of the Invention

The present invention is related to a method for operating a combined multimedia telecom system, more in particular a combined video encodingwireless system, and devices suited therefore.

2. Description of the Related Technology

By discussion of technologies and references in this section, Applicants do not admit that the references are prior art of the invention disclosed in this application.

The current demand in increasing data rate and quality of service in advanced wireless communications has to cope with an energy budget severely constrained by autonomy and ubiquity requirements. Trading off performance and energy consumption deserves the highest attention to enable the ‘anything, anywhere, anytime’ paradigm.

The following observations show the need to integrate the energyefficient approaches across layers. First, stateoftheart wireless systems devices are built to operate at only a fixed set of operating points and assume the worstcase conditions at all times. Irrespective of the link utilization, the highest feasible PHY rate is always used and the power amplifier operates at the maximum transmit power. Indeed, when using nonscalable transceivers, this highest feasible rate results in the smallest duty cycle for the power amplifier. Compared to scalable systems, this results in excessive energy consumption for average channel conditions and average link utilizations. However, recent energyefficient wireless system designs focus on energyefficient VLSI implementations and adaptive physical layer algorithms to adjust the modulation, code rate or transmission power. For these schemes to be practical, they need to be aware of the instantaneous user requirements.

Further, to realize sizeable energy savings, systems need to shutdown the components when inactive. This is achieved only by tightly coupling the MAC to be able to communicate traffic requirements of a single user and schedule shutdown intervals.

In the case of a multiuser wireless communication system, there exist complex tradeoffs between the adaptive physical layer schemes and the requirements of multiple users. For example, lowering the rate of one user affects the available time for the second delay sensitive user. This forces the second user to increase its rate, consume more energy and potentially suffer from a higher bit error rate.

However, the traditional approaches, including most of the stateoftheart crosslayer optimization frameworks, do not yet enable a meaningful tradeoff between performance and energy consumption. Indeed, most of them solve problems in an ad hoc way, focusing on the interaction between adjacent layers and do not raise the scope to the user level. Indeed, the real performance metrics are those quantifying the quality of the service provided by the complete communication stack to the application, while the only effective energy consciousness indicator is the actual energy that is drained from the battery. Both depend jointly on the propagation aspects, the physical layer, the complete protocol stack, the application itself and, more problematically, also on their implementation aspects. This spans far more than the scope of traditional system optimization approaches. Furthermore, the traditional ‘optimization paradigm’ itself, namely finding a unique optimal communication system configuration representing the best tradeoff between performance and cost, becomes inappropriate when the dynamics in the wireless environment and user requirements are considered. More specifically, because of this dynamics, no unique optimal working point exists. The ultimate energyefficient system would have to adapt permanently its characteristic, given the environment constraints, to provide the performance exactly required by the user with the minimum energy.

To achieve this goal flexible systems must be specified having socalled configuration knobs that can be set at runtime to steer jointly the performance and energy consumption. The higher the flexibility, i.e. the number of configuration knobs across all layers, the higher the potential gain due to a better match between the system behavior, the environment and the real user requirements. However, a penalty exists due to the required shift, at least partially, of the optimization process to runtime. This is very challenging due to the combinatorial character of the problem (the number of candidate configurations rises exponentially with the number of controlled knobs).

Recently, joint transmission power and rate control has been considered to reduce system power (see D. Qiao et al., ‘Energy Efficient PCF Operation of IEEE802.11a WLAN with Transmit Power control’, Elsevier Computer Networks (ComNet), vol. 42, no. 1, pp. 3954, May 2003 and ‘MiSer: An Optimal LowEnergy transmission Strategy for IEEE 802.11a/h’, Proc. ACM MobiCom '03, San Diego, September 2003). This approach can be seen as the application to wireless system design of the ‘power aware’ design paradigm proposed by Sinha et al. (‘Energy Scalable System Design’, Trans. on VLSI Systems, April 2002, pp. 135145). Given the fact that transmitting at a lower rate requires less power, the ‘lazy scheduling’ principle has been proposed (see ‘Adaptive Transmission for Energy Efficiency in Wireless Data Networks’, E. UysalBiyikoglu, Ph.D. Thesis, Stanford, June 2003): based on a lookahead of the link utilization (i.e. packet arrival at the transmitter), the minimum average rate to satisfy the current traffic requirements is considered and transmit rate and power are set in function of the channel state in order to achieve this average during the next lookahead window.

In ‘Energyaware Wireless Communications’ (C. Schurgers, Ph.D. thesis, University of California, Los Angeles, 2002) the concept of energyaware radiomanagement is developed. It proposes simple models to capture the energy consumption of radio systems that are used to derive some energyefficient algorithms to select the modulation order, the code rate and to schedule the packets. This dynamic modulation and code scaling is proposed as a practical way to implement lazy scheduling. It also discusses the energy tradeoff between transmission rates and shutting off the system. Operating regions are derived when a transceiver may sleep or use transmission scaling for timeinvariant and timevarying channels. However, the general solutions to transparently exploit the energyperformance scalability at runtime are limited to a few (23) system level knobs. The energyscalability of a system can be defined as the range in which the energy consumption can vary when the performance requirements—e.g. the user data rate—or the external constraints—e.g. the propagation conditions—vary from worst to best case.

In ‘Practical Lazy Scheduling in Sensor Networks’, R. Rao et al, Proc ACM Sensor Networks Conf, Los Angeles, November 2003 a CSMA/CA MAC protocol based on the lazy scheduling idea is derived.

From a theoretical point of view, the ‘lazy scheduling’ concept is attractive. E.g. radio link control based on ‘lazy scheduling’ looks to be a promising technique for WLAN power management. However, it has been analyzed exclusively from the viewpoint of physical, MAC and dynamic link control (DLC) layer. Yet, effective power management in radio communication requires considering the complete protocol stack and its crosslayer interactions.
SUMMARY OF THE INVENTION

Certain embodiments provide a method for operating a telecom system, more in particular a wireless system, with a globally optimized power consumption for a given quality of service. Other embodiments further aim to provide devices suited to the method.

One embodiment includes a method of managing the operation of a system including a processing subsystem configured to run a multimedia application and a telecommunication subsystem. The method includes determining telecom environment conditions, and selecting a working point from a plurality of predetermined working points. The selecting is based at least in part on the determined environmental conditions. The method also includes setting control parameters in the multimedia application and/or the telecommunication subsystem to configure the system to operate at the selected working point, and operating the system at the selected working point.

Another embodiment includes a computer readable medium storing a program configured to execute a method of managing the operation of a system including a processing subsystem configured to run a multimedia application and a telecommunication subsystem. The method includes determining telecom environment conditions, and selecting a working point from a plurality of predetermined working points. The selecting is based at least in part on the determined environmental conditions. The method also includes setting control parameters in the multimedia application and/or the telecommunication subsystem to configure the system to operate at the selected working point, and operating the system at the selected working point.
BRIEF DESCRIPTION OF THE DRAWINGS

The patent or application file contains at least one drawing executed in color. Copies of this paten or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

FIG. 1 represents the energy consumption breakdown in typical wireless transceivers.

FIG. 2A represents DC power consumption of a Class A power amplifier (PA).

FIG. 2B represents adaptation of the PA gain compression characteristic.

FIG. 3 represents the physical layer simulation chain used to derive the PER performance model.

FIG. 4A represents block error rate performance of 8 states of the fading channel realizations.

FIG. 4B is a histogram of the state probabilities of the 8 states of the fading channel realizations.

FIGS. 5A5G are the optimal energy performance tradeoff curves for a 802.11a WLAN link at 10 m, for standard transceivers (dashed line) and energyscalable transceiver (plain line).

FIG. 6 represents a variable rate transmission system model.

FIG. 7 represents the average power versus average delay tradeoff for the optimal (hypothetical) policy (square), the original lazy scheduling (circle) and the proposed proportional control policy (triangle).

FIG. 8 represents the average power vs. average delay for different link load and different values of K. The delay axis is normalized by the control period.

FIGS. 9A9K represent the average delay as a function of the control parameter K in various input bit rates. Up to 80% utilization, the average delay is inversely proportional to K. At higher utilization, the proportionality is lost, the queuing behavior becoming nonlinear due to congestion of the channel.

FIG. 10 represents a wireless pointtopoint file transfer scenario.

FIG. 11 represents the packet drop probability as a function of K.

FIG. 12 represents TCP throughput per connection (i.e. one out of the 10 connections sharing the link) as a function of the control parameter K.

FIG. 13 represents the energy per byte versus TCP throughput tradeoff achievable by tuning the link scheduling policy.

FIG. 14 represents a distributed stackwide power management.

FIG. 15 represents Pareto curves obtained after calibration for a path loss of 80 dB.

FIG. 16 represents the average power versus average goodput tradeoff achieved by the proposed radio link control strategy (circle) compared to the traditional goodput maximizing strategy. The plain line corresponds to the optimal selection.

FIG. 17 represents a centrally controlled LAN/PAN topology illustrating uplink and peertopeer communication.

FIG. 18 represents the determination of a pruned C_{i}(R_{i}) mapping to be used at runtime, starting from a cost and resource profile.

FIG. 19A represents the performance across different channel states.

FIG. 19B is a channel state histogram of a probability of occurrence based on the channel realizations database.

FIG. 20 represents the timing of successful and failed uplink frame transmission under the HCF and the scheme.

FIGS. 21A and 21B represent the mapping for the PA output power and backoff control dimension for a fixed setting of the modulation and code rate control dimensions.

FIG. 22A represents C_{i}(R_{i}) curves in different channel states for 1 fragment.

FIG. 22B represents C_{i}(R_{i}) curves for different frame sizes in channel state 1.

FIG. 23 represents the twoframe buffering.

FIG. 24A represents expected energy consumption across different channel states for 1 fragment frame size.

FIG. 24B represents relative energy consumption by sleeping and scaling for different system loads in the best channel state.

FIG. 25A represents energy consumption per flow as a function of the aggregate system load for CBR traffic.

FIG. 25B represents energy consumption per flow as a function of mean perflow data rate for MPEG traffic.

FIG. 26A represents energy consumption for CBR traffic over a timevariant channel as function of aggregate system load

FIG. 26B represents energy gains for sleeping versus scaling for timevariant channel.

FIG. 27 is a flowchart of an embodiment for managing the operation of a telecom system.

FIGS. 28 to 43 show Pareto points obtained in the EnergyLatency space for DCF and HCF transmission modes.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT

The following detailed description will explain various features of various embodiments in detail. The invention can be embodied in numerous ways as defined and covered by the claims. In order to clarify certain parts of this description some concepts will be illustrated with the practical example of an OFDMbased WLAN system. Some embodiments may be practiced within a QoS management system as described in U.S. Ser. No. 10/254,044, titled “Method for Operating a RealTime Multimedia Terminal and a QoS Manner,” filed Sep. 24, 2002 which is incorporated herein by reference in its entirety. Embodiments herein also relate to EP 1467294, titled “Design Method for Electronic Systems Using Library of Hardware Components with Performance Parameters and Cost Functions,” filed Mar. 25, 2004, U.S. Ser. No. 10/252,688, titled “Method and Device for BlockBased Conditional Motion Compensation,” filed Sep. 19, 2002, and U.S. Ser. No. 10/001,736, titled “Method and Apparatus for Adaptive Encoding Framed Data Sequences,” each of which is incorporated herein by reference in its entirety.

To illustrate certain elements of the solution according to one embodiment, the energyscalability of OFDMbased WLAN such as proposed by the 802.11a/g standards will be considered as an example. Currently proposed techniques to exploit the energyperformance tradeoff in wireless links mainly rely on the reduction of the transmitpower. However, when considering the energy consumption breakdown of practical IEEE 802.11a WLAN transceivers (FIG. 1), it appears that the transmitpower itself accounts only for 8% of the total energy consumption of the transmitter. The Power Amplifier (PA) dominates with 49%. Moreover, due to the stringent linearity and spectral mask requirements of the OFDM modulation, class A power amplifiers are most often required. Consequently, the actual energy consumption of the PA is invariant with the transmit power. On the other hand, traditional receiver architectures are dimensioned for the most demanding modulation and present constant processing gain. Therefore, only the duty cycle effectively affects the link energy consumption. In that context, the highest modulation order that goes through the channel with reasonable packet error rate (i.e. 1%) is always the best choice. Performance (i.e. net data rate) downscaling does not bring effective energy consumption reduction.

In order to get effective energy consumption benefit from energyaware link adaptation technique, the energy consumption scalability of OFDMbased WLAN transceivers has to be enhanced. The following enhancements are proposed: A. configurable power amplifier saturation point; and B. configurable receiver processing gain.

A. Configurable Power Amplifier Saturation Point

The transmitter energyconsumption is dominated by the power amplifier contribution. The idea is to allow adapting the gain compression characteristic of the power amplifier together with its working point on this characteristic. Typical WLAN power amplifiers operate with class A amplifiers that have a fixed gain compression characteristic, so that it is impossible to reduce simultaneously the output power and the linearity, e.g. to adapt to the lower requirement of lower order subcarrier modulations, as illustrated in FIG. 2B. Reducing the output power translates into an increased backoff, and therefore into an higher linearity and a decreased power efficiency. This kills the benefit in terms of total power consumption. Therefore, it makes sense to vary independently the average transmit power and the backoff, which requires to adapt the gain compression characteristic (FIG. 2A). This can be done for instance by applying dynamic voltage scaling.

B. Configurable Receiver Processing Gain

For the receiver, it appears that the energy consumption is dominated by the digital signal processing in which the forward error correction accounts for an important part (FIG. 1). Scalability in receiving energy consumption can be achieved by trading off processing gain (and consequently, receiver sensitivity) and computing energy by activating more or less digital signal processing. This can be achieved for instance by adopting iterative forward error correction (FEC) schemes like turbocodes. Turbocodes can be implemented with low energy consumption and provide a wide freedom to tradeoff energy consumption versus code gain, which translates in receiving processing gain.

The impact of the energyscalability enhancement techniques is now further analyzed. The optimal tradeoff between the net data rate on top of the datalink layer (goodput) and the total link energy consumption is derived by exploring the settings of the traditional link parameters: modulation, code rate, output power, packet size next to the new functional parameters introduced, namely the power amplifier saturation power relative to the output power (backoff) and the number of decoding iterations. Table 1 summarizes the system level control knobs used to bring energyscalability and their range.
 TABLE 1 
 
 
 PA transmit Power (P_{Tx})  0 to 20 dBm by step of 2 dBm 
 PA back off (b)  6 to 16 dB by step of 1 dB 
 Packet (data) size (L_{p})  50, 100, 500, 1000, 1500 bytes 
 Code rate (R_{c})  ½, ⅔, ¾ 
 Modulation order (N_{mod})  1, 2, 4, 6 
 Number of turboiterations (I)  3 to 6 
 

To be able to explore the energyscalability of extended WLAN transceivers, robust performance and energy models are developed, which also consider the specificity of the indoor propagation channel. The models are incorporated in a simulation framework that allows generating very quickly performanceenergy tradeoff curves in realistic user scenarios, e.g. Pareto curves. These models are now discussed more in detail.

Tracking analytically the dependencies between the link parameters (system level knobs), the environment constraints (here the path loss and the channel frequency selectivity), the energy consumption and the performance (data rate) is difficult when the complete system is considered (i.e. not only the physical layer). Performance and energy consumption are nonlinearly coupled e.g. by the automatic repeat request (ARQ) mechanism. To capture accurately those effects, system profiling using eventdriven protocol simulation is most appropriate. Yet, the protocol simulator requires information about the radiolink, i.e. the packet error rate (PER), the gross data rate and the transmission energy per packet, which is also dependent on the system level knobs and the environment constraints.

At the physical layer level performance and energy models can be decoupled. An endtoend simulation chain (FIG. 3) implementing the energyscalability enhancement techniques as described before is used to identify the physical layer PER performance as a function of the system level knobs. First the data is encoded by a turboencoder. Then, the bit stream is modulated using OFDM according to the IEEE 802.11a standard. When transmitted, the signal is hampered by the nonlinearity of the power amplifier. This effect is captured considering an additive distortion power, the distortion being assimilated to a noise source (D). The transmitter output signaltonoise and distortion ratio (SINAD) is related to the backoff according to a law obtained by measurements. Next, the signal passes through the channel. The effect of the wireless indoor channel is decoupled in two effects: first, an average path loss assumed constant (a(d)); secondly, the multipath fading that leads to a timevarying frequency selective channel response. To model the noise introduced by the receiver, a white Gaussian noise (n) is added. Finally, the signal is demodulated and decoded. The Packet error rate (PER) can be evaluated with MonteCarlo simulations. However, it is better to assess the code block error rate (BIER), which is independent of the packet size. The PER can be directly derived from the BIER when knowing the packet size.

Simulation results depict a large variation in performance depending on the channel realization, even when the average path loss is constant. By performing MonteCarlo simulations over a large amount of channel realizations, a classification has been made (FIGS. 4A and 4B). Channel responses leading to similar BIER for the same SINAD have been grouped in representative classes corresponding to socalled channel states. It turns out that 8 classes are sufficient to cover the range with representative average BIER vs. SINAD curves. FIG. 4A depicts the block error rate performance curves for the first 7 states considered. The histogram of the state probabilities is depicted in FIG. 4B. The separation between two adjacent channel classes is set to 2 dB, which corresponds to the granularity with which it is assumed the transmit power can be tuned. No transmission is assumed to be possible in the 8^{th }channel state, which occurs with a probability of less that 0.5%. With such a model, the energyperformance tradeoff can be analyzed independently for each channel state, which again reduces the exploration complexity.

Next to the performance model, an energy model for the wireless link has to be established. The energy to transmit and to receive one packet is given as a function of the system level knobs (Table 2).

For the transmitter, the energy per packet is assumed equal to the product of the radio power by the time needed to transmit the packet (T
_{up}), plus the energy consumed by the digital processing to format it, which is proportional to the packet size. E
_{DSP }denotes the DSP energy consumed per transmitted bit. The radio power is the sum of the PA power (P
_{PA}) and radio frontend power (P
^{T} _{FE}). The PA power (P
_{PA}) is equal to the transmit power (P
_{TX}) divided by the PA power efficiency (η), which is expressed as a function of the backoff b by the measured relation η(b). This relation has been obtained by fitting to measurement results. The contribution of the transmit frontend (P
^{T} _{FE}) is assumed to be constant. The time to transmit a packet (T
_{up}) can be computed as the ratio of the packet size (L
_{p}) by the physical layer data rate, which is the product of the modulation order (N
_{mod}), the number of data carriers in the OFDM symbol (N
_{c}), the code rate (R
_{c}) and the OFDM symbol rate, or baud rate (R
_{s}).
TABLE 2 


Performance model  Energy model 


Model Parameters 
L_{b}  Turbo Code Block size (bit)  288  P^{T} _{FE}  Transmit front  end power (W)  200 
W  Bandwidth (MHz)  20  P^{R} _{FE}  Receive front  end power (W)  200 
T  Temperature (Celsius)  25  E^{R} _{DSP}  Receive DSP energy (nJ/bit)  7.5 
NF  Receiver Noise Figure (dB)  15  E^{T} _{DSP}  Transmit DSP energy (nJ/bit)  8.8 
 (Including Implementation loss)   R_{s}  Baud Rate (kBaud)  250 
d  Distance Transmitter  Receiver (m)  10  N_{c}  Number of data carrier per OFDM symbols  48 
Relations 
A(d)  Channel attenuation  From average propagation  R_{bit}  OFDM bit rate (bps)  R_{bit }= N_{mod }· R_{c }· R_{s} 
  loss model 

D(b)  PA distortion  Fitted on measurement on microsemi PA  T_{up}  Transmit/Receive on time per packet (s)  ${T}_{\mathrm{up}}=\frac{{L}_{p}}{{R}_{\mathrm{int}}}$ 

N  Receiver noise  kT · W · NF  η(b)  PA power efficiency  Fitted on measurement on 
     microsemi PA 

SINAD  Signal to noise and distortion ratio  ${\mathrm{SINAD}}_{c}=\frac{{P}_{\mathrm{tc}}\xb7A}{D\left(b\right)\xb7A+\mathrm{kT}\text{\hspace{1em}}\xb7W\xb7\mathrm{NF}}$  P_{PA}  PA power (W)  ${P}_{\mathrm{PA}}=\frac{{P}_{\mathrm{pTz}}}{\eta \left(b\right)}$ 

BIER  Code Block Error Rate  BIER = f(SINAD, Nmod. Re. 1) Fitted on Monte  Carlo Simulation Results  E_{FEC}  Energy turbo  decoding (J/bit)  $\begin{array}{c}{E}_{\mathrm{FEC}}=3.98\xb7{10}^{9}\xb7{L}_{B}+1.21\xb7{10}^{7}\xb7I+\\ 1.32\xb7{10}^{9}\xb7{L}_{b}+I3.69\xb7{10}^{7}\end{array}\hspace{1em}$ 

PER  Packet Error Rate  $\mathrm{PER}=1{\left(1\mathrm{BIER}\right)}^{\frac{{L}_{p}}{{L}_{b}}}$  E_{TX}  Total Transmit energy (J/bit)  ${E}_{\mathrm{Tx}}=\left({P}_{\mathrm{PA}}+{P}_{\mathrm{FE}}^{T}\right)\xb7{T}_{\mathrm{up}}+{E}_{\mathrm{DSP}}^{T}\xb7{L}_{p}$ 

   E_{RX}  Total Receive Energy (J/bit)  ${E}_{Rx}={P}_{\mathrm{FE}}^{R}\xb7{T}_{\mathrm{up}}+\left({E}_{\mathrm{DSP}}^{R}+{E}_{\mathrm{FEC}}\right)\xb7{L}_{p}$ 


For the receiver, on the other hand, the energy per packet is modeled as equal to the analogue receiver power (P^{R} _{FE}) multiplied by the T_{up}, plus the digital receiver energy per packet, including the turbodecoding energy expressed in function of the number of iterations and the code block size (N). The time to receive the packet is computed the same way as for the transmitter.

Table 2 summarizes the main relations and parameters of the performance and energy models. Parameter values have been captured from measurements carried out on a real advanced WLAN setup.

As performance metrics the user data rate on top of the data link control layer (DLC) is considered and as energy consumption metric, the total energy to transmit successfully a packet. To be able to profile those metrics as a function of the system level knobs settings, considering the impact of the MAC and DLC protocols, the physical layer models are plugged into a network simulator. A twouser scenario has been defined. To evaluate the average user data rate (defined on top of the data link layer) and the corresponding energy consumption, the input queue of the transmitter is filled with 10000 packets and profile the total time and energy consumption needed to receive them correctly in the receiver output FIFO.

The simulation is carried out for every combination of the system level knobs described earlier. This leads to a variety of energyperformance tradeoff points from which only the Paretooptimal ones are retained. The latter points form a socalled optimal tradeoff curve (also called Pareto curve) that gives the minimum energy to achieve a given performance in the considered channel state. Such tradeoff curves are generated for standard and energy scalable systems on the 7 representative channel states. Results are depicted in FIGS. 5A5G. The energy per transmitted bit is considered so that the results are independent of the link duty cycle. For the standard system, the 1dB gain compression point is assumed to correspond to an output power of 36 dBm, thus a backoff of 16 dB for the maximum output power (20 dBm). Standard convolutional codes are considered according to the 802.11a standard. Modulation order, code rate and transmit power are assumed to be adaptive as in stateoftheart link adaptation schemes. Packet size is also assumed to be adaptive, which reveals to have impact only on the worstcase channel states. For the energy scalable system, the additional system level knobs are varied according to the ranges given in Table 1. Both experiments have been carried out assuming an average path loss of 81 dB, corresponding typically to a distance of 10 m. From FIG. 6, it can be seen that with the standard system, it is always better to use the highest modulation order and code rate that go through the channel with reasonable packet error rate (i.e. 1%). This is due to the fact that the increase in energy per bit relative to the usage of a higher order modulation is compensated by the reduction in duty cycle consecutive to the increase in spectral efficiency. In that case, power management techniques based on duty cycle reduction are more effective.

The implication effect of ‘lazy scheduling’ on the endtoend performance of a wireless network will now be analyzed. Further, it is investigated how the endtoend performance versus energy can be effectively controlled by adapting the link layer ‘lazy scheduling’ policy. The performance on top of the transport layer is considered. Indeed, it is the latter that actually provides the communication service to the application. Application specific performance metrics are not considered, but the TCP transport protocol is used, which is by far the most used in the Internet. The impact of variable error rate has already been analyzed (see L. Zou et al., ‘The effects of Adaptive Modulation on the TCP Performance’, Communications, Circuits and Systems and West Sino Expositions, pp. 262266, 2002) but the consequence of the variable rate that would be introduced by ‘lazy scheduling’ has not. TCP throughput degradation resulting from varying rate and delay is discussed in Choi et al. (‘TCP Performance Analysis in Wireless Transmission using AMC’, IEEE VTC Spring, 2003) for CDMA mobile systems using adaptive modulation and coding. However, the possible control of this degradation and the tradeoff with energy are not discussed in any of these related works.

From Shannon one knows that the minimum power P required to reliably transmit on a given (Gaussian) channel (characterized by a given signal to noise ratio and bandwidth) is a monotonously increasing, convex function of the targeted data rate R. This function is given by equation (1) where R_{s }is the symbol rate (baud rate), A and α the constant (average path loss) and variable (fading) components of the channel attenuation. N_{o }is the noise power density.
$\begin{array}{cc}P=\frac{{N}_{0}{R}_{s}}{A\text{\hspace{1em}}\alpha}\xb7\left({2}^{\frac{R}{{R}_{s}}}1\right)& \left(\mathrm{equation}\text{\hspace{1em}}1\right)\end{array}$

When the channel presents a timevarying attenuation, the signal to noise ratio varies accordingly and consequently the feasible rate for a given transmit power. A radio link control scheme is designed that allows finely controlling the link performance versus transceiver power consumption tradeoff by adapting automatically, per frame, the discrete link control knobs introduced previously (modulation, code rate, transmit power and linearity) to tractable channel state information. Adapting the transmit rate and power to timevarying channel conditions in order to maximize the average data rate under an average power constraint is a wellunderstood problem referred to as ‘link adaptation’. Optimal rate and power allocation schemes have been proposed. The dual problem, i.e. minimizing the average power under an average rate constraint can effectively be considered for power management. This can be seen as an application of the ‘power aware’ design paradigm where performance is traded off with energy. From the data link layer point of view (where normally rate and power control are implemented), the performance is traditionally evaluated in terms of net throughput (goodput), which is the net average data rate considering possible retransmissions. When a variable rate is considered, link delay becomes a second important performance metric. Indeed, as shown in FIG. 6, a queue has to be introduced to deal with the transmission rate variation. This queue introduces a delay that dominates the transmission latency at the bottleneck link. It can be shown (Little law) that the average queuing delay ({overscore (D)}) equals the average queue backlog ({overscore (Δ)}) divided by the average rate on the link ({overscore (R)}).

The energy versus queuing delay tradeoff in such systems is extensively studied in ‘Power and Delay Tradeoffs in Fading Channels’, Berry, Ph.D. Thesis, MIT, Cambridge, Mass., December 2000. Longterm average power and queuing latency are considered. Using dynamic programming, policies to specify for each packet, depending on the actual queue backlog and channel gain, which rate to use are characterized. It is shown that queue stability can be guaranteed—i.e. the maximum number of bits in the queue is bounded—and the average energy versus average delay tradeoff is bounded.

‘Lazy scheduling’ is an example of such a policy. The principle of lazy scheduling consists of looking ahead at the packet arrivals, considering a fixed time window (T_{w}). At the end of each time window, the actual queue backlog (Δ_{w}) is considered to compute an average rate constraint (eq. 2), which is used to compute the transmit rate and power as a function of the channel attenuation for the next window. The ‘waterfilling in time’ algorithm can be used for that purpose. Initially, this procedure has been derived considering the information theoretical relation between transmission rate and power (eq. 1). However, this relation corresponds to a bound that practical physical layer modulation and coding schemes tend to approach but do not meet. Also, in practice, the rate cannot be varied continuously but only stepwise, e.g. by varying the modulation order (constellation size) or the coding rate. Without hampering the generality, the practical rate adaptation algorithm proposed by Schurgers (cfr. supra) is considered. In an embodiment, the adaptation policy must indicate which modulation to use as a function of the channel attenuation. It is shown that for a narrow band Rayleigh fading channel, using Quadrature Amplitude Modulation (QAM) with constellation order 2^{b}=2^{2j}, j∈1N, the optimal policy is given by equation (eq. 3), where δ is obtained by solving (eq. 4), R_{s }being the symbol rate of the modulation and {overscore (R)} corresponding to the average rate constraint. The corresponding power is given by (eq. 5) where C is a constant depending on the average path loss, coding gain, receiver noise figure and targeted bit error rate.
$\begin{array}{cc}\stackrel{\_}{R}=\frac{{\Delta}_{w}}{{T}_{w}}& \left(\mathrm{equation}\text{\hspace{1em}}2\right)\\ \begin{array}{cc}{b}_{i}=2\xb7\lceil \frac{1}{2}\xb7{\mathrm{log}}_{2}\left(\frac{{\alpha}_{i}}{\delta}\right)+1& \mathrm{if}\text{\hspace{1em}}{\alpha}_{i}\ge \delta \\ {b}_{i}=0& \mathrm{if}\text{\hspace{1em}}{\alpha}_{i}<\delta \end{array}& \left(\mathrm{equation}\text{\hspace{1em}}3\right)\\ 2.\text{\hspace{1em}}\sum _{j=0}^{\infty}\text{\hspace{1em}}\mathrm{exp}\text{\hspace{1em}}\left(\delta \xb7{4}^{j}\right)=\frac{\stackrel{\_}{R}}{{R}_{s}}& \left(\mathrm{equation}\text{\hspace{1em}}4\right)\\ {P}_{i}=\frac{C}{{\alpha}_{i}}\left({2}^{{b}_{i}}1\right)& \left(\mathrm{equation}\text{\hspace{1em}}5\right)\end{array}$

Notice further that block fading is assumed, i.e. a constant channel attenuation during the transmission of one packet. This requires the channel coherence time to be high relative to the transmit duration of one packet. Hence, the modulation does not have to be adapted during a packet transmission.
 TABLE 3 
 
 
 R_{s}  10 MBaud 
 C  60 mW 
 B  {1, 2, 4, 6} 
 P_{max}  20 dBm 
 _{Tc}  20 ms 
 

Considering this simple but representative model, the energydelay tradeoff achieved by lazy scheduling is evaluated. The channel state is varied according to a 8state finite state Markov model whose transition matrix is set so that the probability density function (pdf) approximates a Rayleigh distribution of average 1 and the coherence time equals T_{c}. Further parameters are summarized in Table 3. The maximum average rate achievable with this setup, considering the transmit power limit of 20 dBm is close to 20 Mbps. In this experience, a constant input bit rate of 10 Mbps is considered, corresponding to 50% utilization of the link capacity.

Results are depicted in FIG. 7 (circles). They are compared to the bound (optimal policy) computed for the same setup (square). Clearly, the initial lazy scheduling performs suboptimally when the tolerated latency is low (a short coherence time). To circumvent this drawback, instead of considering a finite time window, an adaptation of the average rate constraint with a finer grain is proposed. A proportional control is considered. The rate constraint is set proportional to the number of bits in the queue (eq. 6) at a frequency corresponding to the coherence time (T_{c}).
{overscore (R)} _{i} =K×Δ _{i} (equation 6)

Here, {overscore (R)}_{i }and Δ_{i }denote average rate constraint and average delay, respectively, i denoting a time index. The rate/power adaptation policy is tuned continuously according to this varying average rate constraint ({overscore (R)}) in (eq. 4). For a low proportionality factor (K), the system will tolerate a large average delay and react softly to the channel variation. Hence, the adaptation policy will behave close to the unconstrained waterfilling in time. However, for high K, the reaction will be more aggressive, resulting in a lower average delay. For very high K; the system will always tend to select the highest modulation order possible and adapt the power according to channel inversion (i.e. boost the power for bad channels). Hence, by varying K, one can range between the two extreme link adaptation policies: waterfilling in time and channel inversion.

It can be seen from FIG. 7 (triangle) that the new policy beats the initial ‘lazy scheduling’ for smaller latencies and stays close to the hypothetic optimal policy in that region. Powerdelay tradeoffs for different input bit rates are depicted in FIG. 8. It can be seen that our adaptive policy offers a significant tradeoff range for the average rate up to 80% link utilization. For half load (50% utilization), the average power ranges a factor 3 between the extreme adaptation policies. In FIGS. 9A91, the average delay for each of the different input bit rate is plotted as a function of the parameter K. It clearly appears that this knob enables to control the average delay.

Interestingly, one can notice that the proposed policy is separable: the adaptations to the queue backlog—i.e. to the traffic requirements—and to the channel can be decoupled as far as a “constraint propagation” is added. More specifically, an average rate constraint is derived from the queue backlog. This constraint is propagated to tune the rate/power versus channel state adaptation policy. Hence, the solution can be implemented without necessarily jointly designing the rate/power adaptation (in practice, the radio link control layer) and the data link control scheduler (that controls the queue). Those layers can be designed separately, the average rate constraint being the minimal crosslayer information to be passed from one layer to another to guarantee closetooptimal operations.

So far, the energy versus performance tradeoff resulting from the different ‘lazy scheduling’ approaches has been studied from the data link layer viewpoint only. Queuing delay and throughput on a single wireless link are considered as performance metrics. The different algorithms try to minimize the power needed to achieve these given performance metrics. Yet, to effectively study the impact of this technique on the endtoend system performance, it is mandatory to also consider the interactions with the higher layers. Indeed, it should be understood how a bottleneck link (supposed to be the wireless link) delay and throughput translates into endtoend system performance. The endtoend system considered here consists of a single wireless link, corresponding e.g. to point to point scenario (FIG. 10). This is indeed the simplest scenario sufficient to show the crosslayer interactions. The single wireless terminal considered runs a number of applications, e.g. file transfers, which most probably use a TCP/IP protocol stack. The impact of an adaptive link layer on the average endtoend TCP performance will now be shown. First the mechanism of the protocol is briefly reviewed. Secondly, it is shown how average endtoend performance can be controlled by adapting the ‘lazy scheduling’ policy at link level.

TCP offers a reliable connection to the application. To enable this, acknowledgements are used to inform the source if a packet (identified with a sequence number) is well received. Using this feedback mechanism, it is possible to derive implicitly information about the possible network congestion, which occurs when the traffic sent through the network is larger than the capacity of the bottleneck link. Network congestion translates into excessive queuing delays or eventually packets drops at the bottleneck queue. Delays are discovered when the acknowledgement as a response to a sent packet is delayed more than expected by the source (i.e. a timeout event). Packet drops are translated in the reception of ‘tripleduplicate’ acknowledgements, i.e. the acknowledgements of packets following the lost packets contain identical sequence number. TCP reacts on this by maintaining a congestion window of W packets. Each Round Trip Time (RTT), i.e. the time between sending a packet and receiving its acknowledgement, TCP sends W packets. During congestion avoidance, the window is increased by 1/W each time an ACK is received. A TCP connection can also be in the slow start phase, where the window size is increased more aggressively. As one is mainly interested in the steadystate average behavior of TCP, this phase is not considered for the analysis. Conversely, the congestion window is decreased whenever a packets loss is detected, with the amount of the decrease depending on whether packet loss is detected by a duplicate ACK or by a timeout event. For a duplicate ACK, the window size is halved, while for a timeout it is reset to 1.

The steadystate performance of a bulk TCP flow (i.e. a flow with a large amount of data to send, such as file transfers) may be characterized by the send rate, which is the amount of data pushed by the sender per time unit. If the number of packet losses or retransmission is small, the throughput, i.e. the amount of data received per time unit, is well approximated by this send rate. Define p to be the packet loss probability of a sent TCP packet. The send rate of a bulk TCP transfer is well approximated by:
$\begin{array}{cc}S=\frac{1}{\mathrm{RTT}}\sqrt{\frac{3}{2p}}& \left(\mathrm{equation}\text{\hspace{1em}}7\right)\end{array}$

Considering a wireless link with timevarying rate, if the link adaptation is done well and MAC retransmissions are allowed, the losses at the wireless link can be neglected (to e.g. 0.1%). Hence, p is dominated by the losses or delays at the queue. Also RTT is mainly determined by the queuing delays at the bottleneck link (i.e. the wireless link). Therefore, both p and RTT depend largely on the link control parameter K. In FIG. 11, p denotes a range of K. A scenario with 10 TCP bulk transfer connections and a queue size of 128 packets is considered. The packet size is 1500 bytes. Further simulation parameters are the same as in Table 3. The loss probability p is approximated by determining a lower bound of the total number of packets lost or delayed. Hence, this gives a positive approximation of p. For small K, the throughput is determined by the link delay, which is mainly translated into timeout events. A larger K results in more aggressive TCP flows, filling the queue until a drop event, which is translated in more duplicate ACK events. The simulated TCP send rate per connection is plotted in FIG. 13, next to the value computed from equation (eq. 7). The shape of the send rate is indeed well approximated by (eq. 7), although it gives too optimistic values because of the optimistic approximation of p. Based on the link control parameter K, a useful range of throughput values is achieved, resulting in an endtoend energy per byte versus throughput tradeoff as depicted in FIG. 12. For readability, both throughput and energy axis are normalized to their maximum values (165 KB/s and 2 μJ/byte per connection). Energy scales in factor 2.5 when throughput scales in a factor 2, corresponding to K ranging from 2 to 10.

The above analysis shows that the endtoend performance for bulk TCP transfer, i.e. the steadystate endtoend throughput, is mainly determined by the queuing delay, translating into a loss probability p and the round trip time RTT. A performance constraint on top of TCP (throughput) can be translated (by eq. 7) into an average delay constraint on the wireless link, provided that the latter is the bottleneck. When lazy scheduling is considered, this delay can be controlled either by the lookahead window size (original lazy scheduling proposal) or by the proportional control parameter (K) in the scheme according to one embodiment. Also, recall that the proposed link adaptation policy is separable. From the average delay constraints and the actual queue backlog, an average rate constraint can be derived (eq. 2 or 6) and propagated to the radio link control, which can use it to decide, for each packet, which rate and transmit power to use in order to minimize power. This observation has important consequences on the power management design paradigm. Indeed, this shows that efficient power management trading off energy with effective userrelated performance metrics (here the TCP throughput) can be achieved keeping a clean, layered system architecture and its obvious advantage in terms of maintainability. Further, unintended interactions between protocols introduced by flat crosslayer design are avoided. Stackwide power management is achieved by coordinating local algorithms by constraint propagation. The average DLC queuing delay and average transmission rate have been shown to correspond to the minimum information passing required between the transport layer and data link control, and between data link control and radio resource control, respectively. This allows to draft the structure of a stackwide power management scheme distributed between the traditional layers (FIG. 14), still performing as well as a flat crosslayer power management scheme.

Referring to FIG. 14, the layers of the distributed stackwide power management scheme, for example, include the following: application layer, middleware layer, transport layer (TCP), network layer (IP), data link layer (DL) and physical layer (PHY). The data link layer comprises radio link control and logical link control. With the distributed stackwide power management scheme, one can conduct a method of managing the operation of a telecom system which comprises a queuing mechanism introducing a queuing delay. The method comprises: determining telecom environment conditions, which comprises current traffic requirements; determining an average rate constraint to meet the current traffic requirements, by solving an optimization problem and setting at least one control parameter, the control parameter taking into account the instantaneous queuing delay; and operating the telecom system by setting the control parameter taking into account the instantaneous queuing delay. In the method, the average rate constraint is set as a parameterizable function of the number of bits in the queue of the queuing mechanism. The parameters are control parameters. The determination of the average rate constraint is based on a lookahead with a variable window size of link utilization. The window size is determined by said control parameter.

At runtime, the energy and performance model are calibrated (initially and whenever appropriate, still at a very low rate) to derive the actual energy/performance tradeoff characteristics that are used to carry out the runtime settings adaptation at the desired (fast) rate. Parametric energy and performance models being available for the considered system, methods are now derived to carry out the runtime phase. The transmitter configuration is assumed to be done just before the transmission of the packet according to the channel conditions at that moment. The run time phase is split into two steps: a calibration (carried out at low frequency) and an effective runtime adaptation step (per packet). The latter relies on the calibrated energy/performance characteristics. Both steps are analyzed hereafter.

Knowing all parameters, included those measured at runtime (e.g. the average path loss), the energy per bit vs. goodput tradeoff characteristic can be derived for each channel state. From the initial configuration space—i.e. the set of combination of control knobs settings (modulation order, code rate, transmit power and power backoff)—those corresponding to Pareto optimal tradeoffs in the energy per bit versus goodput plane are stored in a table. This corresponds to the socalled calibration. At this point, the energy per bit is considered as energy metrics in order to maintain two independent axes. Indeed, for a given configuration, the energy per bit is constant while the average power depends on the rate and on the duty cycle. If the knobs setting range is limited, an exhaustive search in the configuration space is still possible. If the number of knobs and/or the range is larger, then heuristics should be used. Notice that this search has to be performed only when the model is recalibrated (when entering the network or when the average path loss changes significantly), so the time and energy overhead involved in this is not really that critical. FIG. 15 depicts the results of the calibration with the considered model for an average path loss of 80 dB, corresponding typically to a distance of 10 m indoor. One stepcurve (Pareto curve) represents the Paretooptimal energyrate tradeoff for a given channel state. Each of these points corresponds to a given (N_{mod}, R_{c}, P_{Tx}, b) setting. This implies that selecting one point on one of these curves corresponds to deciding jointly about the carrier modulation (N_{mod}), the code rate (R_{c}), the transmit power (P_{tx}) and the backoff setting (b) to be used.

The knobs settings corresponding to Paretooptimal energyrate tradeoffs are known for each channel state from the calibration. Yet, this is not sufficient to carry out effective energy aware radio link control. A policy to decide which configuration point to use when a given channel state is detected, is needed. A trivial policy would select, on the Pareto curve corresponding to the current channel state, the point providing the goodput just larger than the user data rate requirement. Obviously, such a policy is suboptimal from the power viewpoint. Indeed, since the constraint is the average goodput, it would be more effective to reduce the rate on the bad channel states (where the cost in energy to maintain a given rate is high) and compensate by increasing it on the good channel state (where the cost is lower). This is the principle underlying the waterfilling algorithm proposed in ‘The Capacity of Downlink Fading Channels with Variable Rate and Power’, A. J Goldsmith, IEEE Trans. Veh. Techn., Vol. 46, No. 3, August 1997. Yet, one cannot directly apply this algorithm here due to the discrete nature of the setup. Therefore, first the structure of the problem is analyzed.

Let (r_{ij}, e_{ij}) be the coordinates of the i^{th }Pareto point on the curve corresponding to channel state j. The average power {overscore (P)} and rate {overscore (R)} corresponding to a given radio link control policy—i.e. the selection of one point on each Pareto curve—can be expressed by (eq. 8) and (eq. 9) where x_{ij }is 1 if the corresponding point is selected, 0 otherwise. ψ_{j }is the probability of the channel state j.
$\begin{array}{cc}\begin{array}{c}\stackrel{\_}{P}=\sum _{j}\text{\hspace{1em}}{\psi}_{j}\sum _{i}\text{\hspace{1em}}{x}_{\mathrm{ij}}{e}_{\mathrm{ij}}{r}_{\mathrm{ij}}\\ =\sum _{i}\text{\hspace{1em}}\sum _{j}\text{\hspace{1em}}{x}_{\mathrm{ij}}{\psi}_{j}{e}_{\mathrm{ij}}{r}_{\mathrm{ij}}\stackrel{\u22d2}{=}\sum _{i}\text{\hspace{1em}}\sum _{j}{x}_{\mathrm{ij}}{p}_{\mathrm{ij}}^{\prime}\end{array}& \left(\mathrm{equation}\text{\hspace{1em}}7\right)\\ \begin{array}{c}\stackrel{\_}{R}=\sum _{j}\text{\hspace{1em}}{\psi}_{j}\sum _{i}\text{\hspace{1em}}{x}_{\mathrm{ij}}{r}_{\mathrm{ij}}\\ =\sum _{i}\text{\hspace{1em}}\sum _{j}\text{\hspace{1em}}{x}_{\mathrm{ij}}{\psi}_{j}{r}_{\mathrm{ij}}\stackrel{\u22d2}{=}\sum _{i}\text{\hspace{1em}}\sum _{j}{x}_{\mathrm{ij}}{r}_{\mathrm{ij}}^{\prime}\end{array}& \left(\mathrm{equation}\text{\hspace{1em}}9\right)\end{array}$

The notation p′_{ij }and r′_{ij }is introduced corresponding to the power and rate, respectively, when the channel state is j and the i^{th }point is selected on the corresponding curve, both weighted by the probability to be in that channel state. Only one Pareto point can be selected when being in one channel state, resulting in the following constraints:
$\begin{array}{cc}\sum _{i}\text{\hspace{1em}}{x}_{\mathrm{ij}}=1,\forall j\text{\hspace{1em}}{x}_{\mathrm{ij}}\in \left\{0,1\right\}& \left(\mathrm{equation}\text{\hspace{1em}}10\right)\end{array}$

For a given rate constraint R, the optimal control policy is the solution of the following problem:
$\begin{array}{cc}\mathrm{min}\sum _{i}\text{\hspace{1em}}\sum _{j}{x}_{\mathrm{ij}}{p}_{\mathrm{ij}}^{\prime}\text{\hspace{1em}}\mathrm{subject}\text{\hspace{1em}}\mathrm{to}\text{\hspace{1em}}\sum _{i}\text{\hspace{1em}}\sum _{j}{x}_{\mathrm{ij}}{r}_{\mathrm{ij}}^{\prime}>R& \left(\mathrm{equation}\text{\hspace{1em}}11\right)\end{array}$

This is the classical multiple choice knapsack problem. One is interested in the family of control policies corresponding to R ranging from 0 to Rmax; Rmax being the maximum average rate achievable on the link. This family is called the radio link control strategy. Let kj denote the index of the point selected on the jth Pareto curve. Formally, kj=i
xij=1. A control policy can be represented by the vector k={kj}. The control strategy is denoted {k(n)} corresponding to the set of point {({overscore (R)}
^{(n)},{overscore (P)}
^{(n)})} in the average rate versus average power plane. A good approximation of the optimal radio link control strategy (i.e. that bounds the tradeoff between {overscore (R)} and {overscore (P)}) can be derived iteratively with a greedy heuristic explained in the flowchart shown in Table 4 below. Notice that this derivation is done when the energyperformance tradeoff characteristics per state (the Pareto curves in
FIG. 15) change, i.e. if recalibration occurs.
 TABLE 4 
 
 
 Let k^{(0) }be the solution of (11) when R = 0. Obviously it 
 
 $\mathrm{corresponds}\text{\hspace{1em}}\mathrm{to}\text{\hspace{1em}}\left\{{x}_{0j}=1\text{\hspace{1em}}\forall j\right\},i.e.\text{\hspace{1em}}\mathrm{the}\text{\hspace{1em}}\mathrm{point}\text{\hspace{1em}}\mathrm{of}\text{\hspace{1em}}\mathrm{coordinate}$ 
 (0,0) is selected on each Pareto curve. 
 
 $\mathrm{Denote}\text{\hspace{1em}}\mathrm{the}\text{\hspace{1em}}\text{socalled}\text{\hspace{1em}}\mathrm{current}\text{\hspace{1em}}\mathrm{policy},\text{\hspace{1em}}{k}^{\left(n\right)}=\left\{{k}_{j}^{\left(n\right)}\right\};\mathrm{the}\text{\hspace{1em}}\mathrm{next}$ 
 policy k^{(n+1) }corresponding to the just higher rate 
 constraint can be computed as follows: 
 
 
 $\mathrm{Compute}\text{\hspace{1em}}\mathrm{the}\text{\hspace{1em}}\mathrm{slopes}\text{\hspace{1em}}{s}_{j}^{\left(n\right)}=\frac{{{p}^{\prime}}_{\left({k}_{j}^{\left(n\right)}+1\right)j}{{p}^{\prime}}_{\left({k}_{j}^{\left(n\right)}\right)j}}{{{r}^{\prime}}_{\left({k}_{j}^{\left(n\right)}+1\right)j}{{r}^{\prime}}_{\left({k}_{j}^{\left(n\right)}\right)j}}\forall j$ 
 
 $\mathrm{Define}\text{\hspace{1em}}{\hat{j}}^{\left(n\right)}=\underset{j}{\mathrm{arg}\text{\hspace{1em}}\mathrm{max}}\left({s}_{j}^{\left(n\right)}\right)$ 
 
 $\begin{array}{c}\mathrm{The}\text{\hspace{1em}}\mathrm{next}\text{\hspace{1em}}\mathrm{policy}\text{\hspace{1em}}{k}^{\left(n+1\right)}\text{\hspace{1em}}\mathrm{is}\text{\hspace{1em}}\mathrm{such}\text{\hspace{1em}}\mathrm{as}\text{\hspace{1em}}{k}_{\hat{j}}^{\left(n+1\right)}={k}_{\hat{j}}^{\left(n\right)}+1\text{\hspace{1em}}\mathrm{and}\\ {k}_{j}^{\left(n+1\right)}={k}_{j}^{\left(n\right)},\forall j\ne \hat{j}\end{array}\hspace{1em}$ 
 

From the family of policies derived with the greedy heuristic, it is possible to derive a policy for any average rate constraint ({{overscore (R)}^{(n)}} also). For the given average rate constraint, from the set are selected those policies that lead to a average rate just higher (R_{h}) and just lower (R_{l}) than the constraint (R). For a given packet transmission, the configuration knobs are set according to one of these policies. The selection is done randomly with probability P_{h }and P_{l}=1−P_{h}, respectively, where P_{h }is given by equation (12)
$\begin{array}{cc}{P}_{h}=\frac{R{R}_{l}}{{R}_{h}{R}_{l}}& \left(\mathrm{equation}\text{\hspace{1em}}12\right)\end{array}$

Doing this, the final average power versus average rate characteristic is the linear interpolation between the tradeoff points {({overscore (R)}^{(n)},{overscore (P)}^{(n)})}.

The performance of the proposed radio link controller is evaluated in the average rate versus average power plane. Results are depicted in FIG. 16. The tradeoff curve achieved by the control strategy derived with the greedy heuristic (circles) is compared to the optimal solution of (eq. 11) for the range of R (plain line). Theoretically, the greedy heuristic is suboptimal since the Pareto curves in FIG. 15 are not necessarily convex. Yet, it clearly appears that the optimal solution is closely approached. This tradeoff curve can be used to evaluate the benefit of the proposed strategy compared with traditional radio link control aiming at maximizing the data rate under power constraint. With such a strategy, the policy corresponding to the highest feasible rate would be selected. This corresponds to the point A in FIG. 16. If the link capacity is not fully used, the link idles and the terminal enters regularly in sleep mode. The average power decreases linearly with the duty cycle. Hence, the corresponding powerrate tradeoff is represented by the dasheddot line in FIG. 16. It can be seen that, in all cases, except the one corresponding to maximal load, the proposed energy aware radio link control strategy saves power when compared with the traditional strategy. For example at half load, the gain reaches 80%.

Another embodiment relates to wireless communication systems that are provided with a sleep mode. Further also the multiuser aspect is introduced in the scheme to manage systemwide power management dimensions at runtime as described below.

In the wireless network as in FIG. 17 multiple nodes are controlled centrally by an access point (AP). Each node (such as e.g. a handheld video camera) wishes to transmit frames at realtime and it is the AP's responsibility to assign channel access grants. In an embodiment, the resource allocation scheme within the AP must ensure that the nodes meet their performance constraints by delivering their data in a timely manner while consuming minimal energy. More formally the problem can be stated in the following way. The network consists of n flows {F_{1}, F_{2}, . . . , F_{n}}. For notational simplicity, a onetoone mapping of flows to nodes is assumed, but the design methodology is applicable to one or more flows per node. Each flow i, 1≦i≦n, is described by the following properties:

(a) Cost Function (C_{i}): This is the optimization objective, e.g. to minimize the total energy consumption of all users in terms of Joules/job. In for example a video context, a job is the frame size of current application layer video frame.

(b) QoS Constraint (Q_{i}): The optimization has to be carried out taking into account a minimum performance or QoS requirement in order to satisfy the user. As delivery of realtime traffic is of interest (e.g. video streaming), the QoS is described in terms of the job failure rate (JFR) or deadline miss rate. JFR is defined as the ratio of the number of frames not successfully delivered before their deadline to the total number of frames issued by the application. The QoS constraint is specified by the user as a targetJFR (JFR*), to be maintained over the lifetime of the flow.

(c) Shared Resource {R_{1}, R_{2}, . . . , R_{m}}: Multiple resource dimensions could be used to schedule flows or tasks in the network, e.g. time, frequency or space. The restricted case is considered here where access to the channel is only divided in time. Therefore, time is the single shared resource and is denoted by R. The resource fraction consumed by the i^{th }node is denoted by R_{i}. The maximum time available for any flow is R_{i} ^{max}, which is the frame period for periodic traffic.

(d) Control Dimensions {K_{1}, K_{2}, . . . , K_{l}}: For a given wireless LAN architecture, there are platform independent control knobs or dimensions as already described before that control the received signaltonoise ratio related to the resource utilization in terms of the transmission time per bit, given the current path loss. The control dimension settings are discrete, interdependent and together have a nonlinear influence on the cost function.

(e) System state {S_{1}, S_{2}, . . . , S_{s}}: As the environment is very dynamic, the system behavior will vary over time. The environmental factors independent of the user or system control are represented by a system state variable. In a wireless environment with e.g. VBR video traffic, the system state is determined by the current channel state and the application data requirements. The scheduling algorithm is executed periodically based on the channel epoch and the rate at which the data requirements change. Each flow is associated with a set of possible states, which determines the mapping of the control dimensions to the cost and resource.

The following system properties contribute to the structure of the methodology. The key aspects are the mapping of the control dimensions to cost and resource profiles respectively, and the general properties of this mapping. A resource (cost) profile describes a list of potential resource (cost) allocation schemes needed for each configuration point K. These profiles are then combined to give a CostResource tradeoff function, which may be essential for solving the resource allocation problem.

Cost Profile Properties

 Every flow has a known minimum and maximum cost over all control dimensions, which is a function of the desired JFR* and the current system state (e.g. channel state). The cost range (difference between maximum and minimum) is important as this determines the impact the control dimensions have on the system cost.
 The perdimension discrete control settings can be ordered according to their minimal associated Cost.
 The overall system cost, C, may be defined as the weighted sum of costs of all flows, where each flow can be assigned a certain weight depending on its relative importance or to improve fairness.
$\begin{array}{cc}C=\sum _{i=1}^{n}\text{\hspace{1em}}{w}_{i}{C}_{i}& \left(\mathrm{eq}.\text{\hspace{1em}}13\right)\end{array}$

Resource Profile Properties

 Every flow has a known minimum and maximum resource requirement across dimensions, which is a function of the desired JFR* and system state.
 Depending on the current system constraints and possible configurations, each flow has a minimum resource requirement R_{i} ^{min}. It is assumed all minimum resource requirements can be satisfied at each moment in time. Hence, no overload occurs and all flows can be scheduled. This is a reasonable assumption under average load and average channel conditions. However, under worstcase conditions and nonscalable video applications, a system overload may occur and one or more flows will need to be dropped. While the policy to drop flows is out of the scope of the optimization criterion, a practical system may employ a policing policy that is fair to the users.
 The perdimension discrete control settings can be ordered according to their minimal associated Resource requirement.
 The overall system resource requirement R is defined as the sum of the per flow requirements:
$\begin{array}{cc}R=\sum _{i=1}^{n}\text{\hspace{1em}}{R}_{i}& \left(\mathrm{eq}.\text{\hspace{1em}}14\right)\end{array}$

The goal is to assign transmission grants, resulting from an optimal setting of the control dimensions, to each node such that the per flow QoS constraints for multiple users are met with minimal energy consumption. For a given set of resources, control dimensions and QoS constraints, the scheduling objective is formally stated as:
$\begin{array}{cc}\underset{C}{\mathrm{min}}\sum _{\mathrm{nodes}}\text{\hspace{1em}}{\omega}_{i}{C}_{i}\left({S}_{i}\right)\text{}\mathrm{subject}\text{\hspace{1em}}\mathrm{to}\text{:}\text{}\begin{array}{cc}{\mathrm{JFR}}_{i}\le {\mathrm{JFR}}_{i}^{*},i=1,\dots \text{\hspace{1em}},n& \left(\mathrm{QoS}\text{\hspace{1em}}\mathrm{Constraints}\right)\\ \sum _{\mathrm{nodes}}{R}_{i,j}\le {R}_{j}^{\mathrm{max}},j=1,\dots \text{\hspace{1em}},m& \left(\mathrm{Resource}\text{\hspace{1em}}\mathrm{Constraints}\right)\\ {K}_{j}>{R}_{i}\left({S}_{i}\right),i=1,\dots \text{\hspace{1em}},n;\text{\hspace{1em}}j=1,\dots \text{\hspace{1em}},l& \left(\mathrm{Resource}\text{\hspace{1em}}\mathrm{Profiles}\right)\\ {K}_{j}>{C}_{i}\left({S}_{i}\right),i=1,\dots \text{\hspace{1em}},n;\text{\hspace{1em}}j=1,\dots \text{\hspace{1em}},l& \left(\mathrm{Cost}\text{\hspace{1em}}\mathrm{Profiles}\right)\end{array}& \left(\mathrm{eq}.\text{\hspace{1em}}15\right)\end{array}$

The solution of the optimization problem yields a set of feasible operating points K, which fulfill the QoS target, respects the shared resource constraint and minimizes the system cost.

When considering energyscalable systems, the number of control dimensions is large and leads to a combinatorial explosion of the possible system configurations as already explained. In addition, the resource and cost profile relations are complex. In order to solve this problem efficiently, a pragmatic scheme is needed to select the configurations at runtime. This is achieved by first determining the optimal configurations of all control dimensions at design time. At runtime, based on the channel condition and application load, the best operating point is selected from a significantly reduced set of possibilities.

A property of the designtime phase model is that the configurations can be ordered according to their minimal cost and resource consumption, describing a range of possible costs and resources for the system. For each additional unit of resource allocated, one only needs to consider the configuration that achieves the minimal cost for that unit of the resource. For each possible system state (for different channel and application loads), the optimal operating points are determined by pruning the CostResource (CR) curves to yield only the minimum cost configurations at each resource allocation point.

A function p_{i}: R→C is defined, such that
p _{i}(R _{i}(S _{i}))=min{C _{i}(S _{i})(K _{i} →R _{i}(S _{i}))ˆ(K _{i} →C _{i}(S _{i}))}
which defines a mapping between the Resource and the Cost of a certain configuration, k, for a node in a state, S_{i}, as shown in FIG. 18. Considering the resulting points in the CR space, one is only interested in those that represent the optimal tradeoff between the energy and resource needs for the system. Indeed, the tradeoff between transmission time and transmission energy is a fundamental property for wireless communication, bounded by Shannon. Although the discrete settings and nonlinear interactions in real systems lead to a deviation from this optimal tradeoff, it can be well approximated as follows.

The convex minorant of these pruned curves is calculated along both for the Cost and Resource dimensions, and the intersection of the result is considered. As a result, the number of operating points is reduced significantly thus rendering it very useful for the runtime optimization (FIG. 18).

Several tradeoffs are present in the system: increasing the modulation constellation size decreases the transmission time but results in a higher packet error rate (PER) for the same channel conditions and PA settings. In an embodiment, the energy savings due to decreased transmission time must offset the increased expected cost of retransmissions. Also, increasing the transmit power increases the signal distortion due to the PA. On the other hand, decreasing the transmission power also decreases the efficiency of the PA. Similarly, it is not straightforward when using a higher coding gain, if the decreased SNR requirement or increased transmission time dominates the energy consumption. Finally, considering the tradeoff between sleeping and scaling: a longer transmission at a lower and more robust modulation rate needs to compensate for the opportunity cost of not sleeping earlier.

A greedy algorithm is employed to determine the perflow resource usage R_{i }for each application to minimize the total system cost C. The algorithm traverses all flows' CostResource curves and at every step consumes resources corresponding to the maximum negative slope across all flows. This ensures that for every additional unit of resources consumed, the additional cost saving is maximum across all flows. The current channel state and application demand are assumed to be known for each node. This information is obtained by coupling the MAC protocol with the resource manager and is explained in the next section. By the assumption that sufficient resources are available for all flows, the optimal additional allocation to each flow, R_{i}>0, 1≦i≦n, subject to Σ_{i=1} ^{n }R_{i}≦R is determined.

The following greedy algorithm is used:

a. Allocate to each flow the smallest resource possible for the given state, R_{min}. All flows are schedulable under worstcase conditions, hence Σ_{i=1} ^{n }R_{min}≦R.

b. Let the current normalized allocation of the resource to flow, F_{i}, be R_{i}, 1≦i≦n. Let the unallocated quantity of the available resource be R_{l}.

c. Identify the flow with the maximum negative slope, C_{i}′(R_{i})—representing the maximum decrease in cost per resource unit. If there is more than one, pick one randomly. If the value of the minimum slope is 0, then stop. No further allocation will decrease the system cost further.

d. Increase R_{i }by the amount till the slope changes for the i^{th }flow. Decrement R_{l }by the additional allocated resource and increment the cost C by the consequent additional cost. Return to step b until all resources have been optimally allocated or when R_{l }is 0.

In this implementation, the configuration points at designtime are sorted in the decreasing order of the negative slope between two adjacent points. The complexity of the runtime algorithm is O(n.log L) for n nodes and L configuration points per curve. For a given channel and frame size, the number of configuration points to be considered at runtime is relatively small.

Taking into account that the relation C_{i}(R_{i}) derived at design time is a convex tradeoff curve, it can be shown easily that the greedy algorithm leads to the optimal solution for continuous resource allocation. The proof can be extended for real systems with discrete working points to show that the solution is within bounded deviation from the optimal. For a real system, however, the settings for different control dimensions such as modulation or transmit power are in discrete units. This results in a deviation from the optimal resource assignment. The worstcase deviation from the optimal strategy is bounded and small.

Again the practical example of an OFDMbased wireless LAN system is now considered. As previously discussed, several control dimensions can be identified that enable to tradeoff performance for energy savings and vice versa. As above the power amplifier backoff (P_{backoff}), the power amplifier transmit power (P_{TX}), the modulation (N_{Mod}) and the code rate (B_{c}) are considered.

To determine the Job Failure Rate and total expected energy consumption, it may be essential to consider the system dynamics, which include the current channel condition and the application demand. The current job size (in number of fragments) varies significantly for video traffic. In addition, just considering the average received SINAD is not sufficient to characterize the channel for wireless OFDM systems where frequencyselective fading is important. A timevarying channel model is considered and expressions are derived relating the average packet error rate (PER), the JFR and expected energy consumption.

1) Traffic Model

As the goal is to provide timeliness (QoS) guarantees while minimizing energy consumption for a target performance, periodic delaysensitive traffic is considered. Both constant bit rate (CBR) and variable bit rate (VBR) traffic is studied, in order to show the impact of the dynamics. VBR traffic consists of MPEG4 flows. A Transform Expand Sample (TES) based MPEG4 traffic generator that generates traffic with the same first and second order statistics as an original MPEG4 trace is used. All fragmentation is done at the link layer and if a frame is not completely delivered to the receiver by its deadline, it is dropped. All applications employ UDP over IP.

Each frame size in fact maps to a different system state. A frame size is determined in a number of MAC layer fragments, which is assumed to be 1024 bytes long for this experiment. From the results, it is observed that for a given frame size, extrapolating the results for a curve within five fragments results in a very low approximation error. As the maximum frame size is assumed to be 50 fragments long in the tests considered, CostResource curves are only constructed for 1, 2, 3, 4, 5, 10, 20, 30, 40, 50 fragments per frame.

2) Channel Model

A frequency selective and time varying channel model is now used to compute the PER for all transceiver knob settings. An indoor channel model based on HIPERLAN/2 was used for a terminal moving uniformly at speeds between 0 to 5.2 km/h (walking speed). This corresponds theoretically to a coherence time of approximately 28 ms. A set of 1000 timevarying frequency channel response realizations (sampled every 2 ms over one minute) were generated and normalized in power. Data was encoded using a turbo coder model and the bit stream was modulated using 802.11a OFDM specifications. For a given backoff and transmit power, the SINAD at the receiver antenna was computed as before. A pathloss of 80 dB at a distance of 10 m is assumed.

The signal was then equalized (zeroforcing scheme), demodulated and decoded. From the channel realization database, a onetoone mapping of SINAD to receive block error rate was determined for each modulation and code rate. The channel was then classified into 5 classes (FIG. 19A). In order to derive a timevarying linklayer error model, we associate each channel class to a Markov state, each with a probability of occurrence based on the channel realizations database (FIG. 19B). Given this fivestate error model, the PER can be efficiently modeled for different configurations at runtime. The Packet Error Rate (PER) is obtained by assuming the block errors follow a binomial process for a packet size of L_{frag }bits and a block size of 288 bits (see also Table 2):
PER=[1−(1−BlER)^{L} ^{ frag } ^{/288}] (eq. 16)

Now the exact mapping of the control dimensions K to the cost and resource dimensions is derived, based on these expressions and the system state. When delivering a frame, different transmission and retransmission strategies can be used, each resulting in a different total expected energy and time to send the frame, and each resulting in another expected failure rate for the frame. To simplify, the policy is adopted that each fragment of a frame should be transmitted or retransmitted using the same configuration K. This is a good approximation for the real optimal transmission strategy, which includes adapting the strategy depending on the outcome of a fragment transmission (conditional recursion which is complex to solve). For the approximation, a recursive formulation can be derived to compute the expected energy E_{K}, the timeslot needed TXOP_{K}, and the expected failure rate JFR_{K}, for each system state. The MAC protocol overhead is taken into account in this mapping, and the parameters, which are based on 802.11e, are listed in Table 2. A Contention Free access scheme is considered for the transmissions.

Consider a frame, which consists of m fragments or packets and has to be delivered during a known channel state C_{S}. The tuple (C_{S},m) is defined as the system state. All following expressions for cost and resource consumption, hence, depend not only on the configuration K, but also on the system state. For notational simplicity, the state index is omitted. Each packet is transmitted with configuration K, for which the PERK can be determined, based on the models derived above. The probability that the frame is delivered successfully with exactly (m+n) transmissions (hence n retransmissions), is given by the recursion:
$\begin{array}{cc}{S}_{n}^{m}\left(K\right)=\sum _{i=1}^{\mathrm{min}\left(m,n\right)}{C}_{i}^{m}\times {\left({\mathrm{PER}}_{K}\right)}^{i}\times {\left(1{\mathrm{PER}}_{K}\right)}^{mi}\times {S}_{ni}^{i}\times {S}_{ni}^{i}\left(K\right)& \left(\mathrm{eq}.\text{\hspace{1em}}17\right)\\ {S}_{0}^{m}\left(K\right)={\left(1{\mathrm{PER}}_{K}\right)}^{m}& \left(\mathrm{eq}.\text{\hspace{1em}}18\right)\end{array}$
in which C_{i} ^{m }denotes the number of possibilities to select i objects out of m.
Hence, the probability to deliver the frame consisting of m fragments correctly with maximum n retransmissions is
$\begin{array}{cc}1{\mathrm{JFR}}_{n}^{m}\left(K\right)=\sum _{j=0}^{n}{S}_{j}^{m}\left(K\right)& \left(\mathrm{eq}.\text{\hspace{1em}}19\right)\end{array}$
Here only data losses are assumed to result in job failures. As control frames are much shorter and less susceptible to errors, it is assumed they do not suffer from packet errors.

In order to determine the expected energy and time needed to deliver a frame with m fragments and n retransmissions, one needs to know the overhead due to the MAC protocol for a successful and a failed transmission, i.e. E_{good}, E_{bad}, T_{good }and T_{bad}. As a 802.11e HCFtype MAC is assumed, a successful and failed data transmission follow FIG. 20. Let E_{ACK }be the energy needed to receive an ACK packet, and T_{ACK }the time needed for the ACK, E_{Header }and T_{Header }are the overheads for the MAC and PHY headers.
E _{good}(K)=E _{K} +E _{Header}+(2×T _{sifs} ×P _{Idle})+E _{ACK} (eq. 20)
E _{bad}(K)=E _{K} +E _{Header}+((2×T _{sifs} +T _{ACK})×P_{Idle}) (eq. 21)
T _{good}(K)=T _{K} +T _{Header}+(2×T _{sifs})+T _{ACK} (eq. 22)
T _{bad}(K)=T _{good}(K) (eq. 23)
The time needed to send m fragments with max n retransmissions, for configuration K, is then:
TXOP _{n} ^{m}(K)=[m×T _{good}(K)]+[n×T _{bad}(K)] (eq. 24)

The average energy needed to transmit m fragments, with maximum n retransmissions, and configuration K is a bit more complex. The reason is that one is interested in the expected energy, taking into account the chance that a retransmission should happen or not:
$\begin{array}{cc}{E}_{n}^{m}\left(K\right)=\sum _{j=0}^{n}{S}_{n}^{m}\left(K\right)\times \left(\left(m\times {E}_{\mathrm{good}}\left(K\right)\right)+\left(j\times {E}_{\mathrm{bad}}\left(K\right)\right)\right)& \left(\mathrm{eq}.\text{\hspace{1em}}25\right)\end{array}$

It is the sum of the probability that the transmission will succeed after m good and j bad transmissions, times the energy needed for these good and bad transmissions. In order to have the correct expected energy consumption, a second term should be added to denote the energy consumption for a failed job, hence when there are less than m good transmissions, and (j+1) bad ones:
$\begin{array}{cc}{E}_{n}^{m}\left(K\right)={E}_{n}^{m}\left(K\right)+{\mathrm{JFR}}_{n}^{m}\left(K\right)\times \hspace{1em}\left[{E}_{\mathrm{bad}}\left(K\right)+\sum _{j=1}^{m}{S}_{n}^{j}\left(K\right)\times \left(\left(j\times {E}_{\mathrm{good}}\left(K\right)\right)+\left(n\times {E}_{\mathrm{bad}}\left(K\right)\right)\right)\right]& \left(\mathrm{eq}.\text{\hspace{1em}}26\right)\end{array}$

As a result, the E, TXOP, and JFR can be determined as a function of frame size, channel state and number of retransmissions for each configuration K. This determines the full cost and resource profile for the system. In FIGS. 21A and 21B, the impact of the PA control knobs (PA backoff and PA transmit power) on the resource (TXOP) and cost (energy) is illustrated. Only the mapping that corresponds to the smallest TXOP and energy consumption is plotted (hence the mapping with the smallest number of retransmissions that achieves the required JFR* of 10e3). FIGS. 22A and 22B shows the merged and pruned energyTXOP curves for different channel states and different frame sizes, respectively. One can see that the total range in energy consumption is large, both within and across system states. The large tradeoff illustrates the fact that traditional systems, which are designed for a fixed and worst cast scenario, result in a significant energy waste.

Determining a schedule that achieves the required performance with minimal energy consumption is challenging because the instantaneous load and channel state vary independently for each node. Previously the Energy and TXOP were determined needed to deliver a job depending on the current state, which is the frame size and the channel state. Based on these curves for each node, the scheduler can derive a nearoptimal allocation at runtime. There needs to be a feedback loop between the nodes and the scheduler in the AP. The current state information needs to be collected by the scheduler and the decisions about the channel access grants should be communicated back to the nodes with minimal overhead. It is now shown how a sleepaware Medium Access Controller can take care of this.

The MAC is responsible for resource allocation of the shared channel among different users. The packetscheduling algorithm in the AP decides which node is to transmit, when, and for how long. The focus of frame scheduling is on the particular cases of frames sent from a node to its associated AP (uplink) and also from one node directly to another and not via the commonly associated AP (peertopeer), as in FIG. 17. As all nodes are assumed to be within the transmission range of the AP, the AP can only schedule one peertopeer communication simultaneously. All nodes communicate only when they have a valid channel access grant or transmit opportunity (TXOP) or timeslot from the AP.

In order to instruct a node to sleep for a particular duration, the AP needs to know when the next packet must be scheduled. Waking a node earlier than the schedule instance will cause it to waste energy in the idle state. Waking the node later than the schedule instance, will cause it to miss the packet's deadline or waste system resources. The sleepaware MAC protocol therefore employs two techniques to eliminate data dependency due to the application and channel. By buffering one frame, the AP can instruct the node to sleep from the time it was informed that the frame arrived to its scheduling instance. The AP still needs to poll the node upon frame arrival and therefore this only permits the node to sleep between the packet arrival instance and the packet schedule instance.

Buffering just two frames informs the AP of the current traffic demand but also the demand in the next scheduling instance. As shown in FIG. 23, the AP now only needs to communicate with the node at scheduling instances. As the realtime stream's packets are periodic, all idle time between transmission instances is eliminated for the ideal scheme.

In order to inform the AP of the instantaneous channel and the required application load for the current and next scheduling instances, one needs to employ a linklayer feedback mechanism. This is accomplished by adding just three bytes in the MAC header for the current channel state and the two buffered frame sizes. Protocols such as 802.11e provide support for channel information and queue sizes therefore require only minor modifications. In every transmission to the AP, the node communicates its channel state and packet sizes of the two heads of the line packets. In the acknowledgement, the AP instructs the node to sleep until the time of the next scheduling instance and also assigns it the duration of its next TXOP or resource slot. The scheduling decision is, thus, made every frame period (e.g. 30 ms for highquality video) of the flow in the system with the highest frame rate.

The energy savings are now verified over a range of practical scenarios. For all results presented here, the target JFR* is set to 10^{−3 }which is a reasonable value for wireless links. First the expected energy savings are analyzed across all channel states and the entire range of system loads. While this is profiled at designtime, it provides insight as to (a) the range and variation of energy savings with the system dynamics and (b) the contributions of energy savings from sleeping and scaling under different conditions. Following this, the energy savings at runtime are studied for both constant bit rate (CBR) and MPEG4 video flows under different channel conditions and different system loads. In order to evaluate the relative performance of MEERA, the following comparative transmission strategies are considered:

 1. MEERA: This is the optimal operating scheme considering the energy tradeoff between sleep and scaling. The operating point is determined from the CR curves and the runtime algorithm described previously.
 2. MEERAno sleep: This scheme uses the CR curves to determine the optimal TXOP when no sleeping is possible. The same runtime algorithm is used and the nodes remain in the idle state after completion. The purpose of this case is to show the contribution of sleeping under different loads.
 3. Fixed: In this scheme the backoff and output power are fixed to the highest setting and use take the highest feasible modulation and code rate that will successfully deliver the packets. After successful transmission, one goes to sleep. This case is the approach where no scaling is employed and the sleep duration is maximized.
 4. Fixed—no sleep: This scheme is similar to Fixed, but the transceiver remains in the idle mode after successful transmission. This is the base operating scheme of current wireless LAN transceivers with no power save features enabled.

Consider the CR curves in FIG. 22, for transmitting a onefragment sized frame over different channel states. If the TXOP assignment for the user is distributed uniformly over time, the expected energy consumption is proportional to the area under the C_{i}(R_{i}) curve for each channel state shown in FIG. 22(a). Similarly, the CR curves are derived for the three comparative transmission strategies discussed above. The C_{i}(R_{i}) curve, for the fixed case, for example, is in fact a fixed point (at small TXOP), and the expected energy consumption proportional to the area under a horizontal line through this point. In FIG. 24A, the relative energy consumption (normalized by the maximum energy consumed by Fixed over all cases), is plotted for the four schemes over different channel states. As expected, MEERA outperforms the other techniques since it takes advantage of the energy that can be saved by both sleeping and TXOP scaling. The energy needed to transmit a unit of data increases from best to worst channel state due to a combination of (a) the lower modulation rate necessary to meet the higher SINAD requirement (hence shorter sleep duration), (b) a higher required output power to account for the worse channel and (c) the increased cost of retransmissions. It can be observed, for example, for the best channel state, the energy consumption is low for both the Fixed and MEERA approaches. The energy gains for this channel state primarily result from sleeping. On the other hand, for the worst channel state, the transmission energy becomes more dominant, the energy gains due to TXOP scaling are more significant. When looking at the energy gains contributed by sleeping and scaling over a range of link utilizations, the energy gains due to TXOP scaling can be expected to dominate when the transmission duration is large when compared to the sleep interval. Hence, for larger frame sizes or at higher perflow link utilization, the relative energy saving due to scaling has a grater influence. This observation is illustrated in FIG. 24B, where the relative gain for the different techniques—compared to the Fixedno sleep case—are plotted over a series of frame sizes (in terms of number of 1024 byte fragments), for channel state 1. Indeed, for a 5fragment frame TXOP, scaling and sleeping (represented by MEERANo sleep and Fixedsleeping respectively) have a comparable performance. For larger frame sizes, the TXOP scaling in MEERANo sleep contributes significantly to the energy savings.

Now a multipleuser scenario is considered where the TXOP assignments are not uniformly distributed but based on the user's application data rate requirement and the constraints enforced by other users sharing the same link. The influence is now discussed of the aggregate link utilization on the perflow energy consumption for CBR flows over a static channel. The effective throughput of 802.11e, after considering protocol overheads for the first channel state, is approximately 30 Mbps when the highest modulation constellation is used. In the experiment illustrated by FIG. 25A, the link utilization is increased in steps of 2 Mbps CBR flows up to the maximum link capacity. The perflow energy consumption of MEERA increases as the aggregate system load increases. At higher loads due to a large number of flows, a smaller TXOP from the CR curve is assigned to each flow resulting in higher per flow energy consumption. The difference with the case where only TXOP scaling is used (MEERAno sleep) is most noticeable since the possibility to scale is reduced with increasing system load. In a multiuser scenario it is always beneficial to include sleeping as it is influenced to a lesser extent by aggregate load increases due to other flows. On the other hand, if the load is increased by increasing the perflow throughput requirement for a fixed number of flows, the gain of scaling dominates over sleeping. In FIG. 25B, the average perflow data rate requirement is increased for an MPEG4 flow (in a network with five such flows). In this case, the reduction of sleep duration increases the energy consumption only slightly as the energy saving from scaling start contributing more. This is evident when comparing MEERA with the energy consumption of Fixed, which increases almost linearly with the data rate. It is important to note that as the system is not forced into overload, only moderate link utilization is considered (<70% average link utilization using the highest transmission rate) for MPEG4 flows.

Now the energy consumption trends are considered for a timevarying channel. A 5user scenario is used to understand the impact of dynamic channel variations on energy consumption. The channel varies independently over all the users on a framebyframe basis. In FIG. 26A, the total system load is increased from 2.5 Mbps to 10 Mbps for five CBR flows. When compared to the same system load where the channel is in the best (fixed) state, an increase in energy consumption is observed. This is because during every scheduling cycle, the flows experiencing worse channel states require more transmission time (due to lower constellation) and therefore consume more energy. In addition, they force the other flows to transmit in a smaller TXOP and increase their energy consumption too. Further, in FIG. 26B the energy savings achieved when compared to Fixedno sleep over different system load are shown. Sleeping results in an energy saving for all loads, which is obvious in a multiuser scenario. However, comparing the gain MEERA achieves to the Fixed approach (with sleeping), MEERA results in a significant additional gain of a factor 2 to 5 depending on the system load. The combination of sleep and scaling in MEERA yields an overall system gain factor from 2 to 9.

FIG. 27 is a flowchart of a process for managing the operation of a telecom system in accordance with another embodiment. According to the embodiment, the process comprises determining a rate constraint and a telecom environment condition (S2701) and providing predetermined working points per discrete set of telecom environment condition (S2703). The process may further include determining the predetermined working points. Then, a working point is selected from the predetermined working points (S2705), and the telecom system is operated at the selected working point (S2707). Prior to the selection of the working point, the process may further include loading and adapting the predetermined working points.

The optimization process can lead to much better solutions if all the factors are taken into account in a crosslayer approach. This way, we explore the crosslayer interdependencies and joint optimizations for varying network and video coding settings.

In one embodiment the application of this approach is demonstrated for a single user, singlehop wireless link, where the transmission occurs from the AP (Access Point) to the mobile device. But the generic concept is applicable directly in multiuser TDMA based contexts as well. In an exemplary embodiment we have also based the entire demonstration on industryaccepted standards for the layers, i.e. MPEG4 main profile, UDP, IP (Internet), WLAN 11a/e with a DCF or HCF scheduler. Other standards may also be used.

A main contribution is the identification of important latencyenergy tradeoffs that can be exploited for any multi or singleuser service involving interactive passing of data. In that case the user(s) can exchange a longer duration of their battery with a larger response time for the interaction over the wireless network (including the application layer). These tradeoffs are only exposed through a full crosslayer modeling and analysis of the entire protocol stack. We have also introduced several runtime controllable knobs to enable this tradeoff.

Some embodiments deal with batterypowered multimedia (in particular a video encoding system) combined with a telecommunication system (or services) (in particular a wireless system, with either single or multiple users, with direct (singlehop) or indirect (multiplehop) communication. Some embodiments provide an energylatency management system (and even energylatency—Quality management system) for such combined system, said management system including storage of predetermined latencyenergy tradeoff's, where from said management system selects on the fly (runtime) a working point.

In one embodiment the settings for the video codec are fixed for the whole sequence.

In another embodiment an online video adaptation to the network condition is used.

Some embodiments have multiple predetermined operating points, classified to external conditions (for instance the channel state), said predetermination being done by performing optimization across the various layers, meaning by exploring at least a parameter from the multimedia application and a parameter from the wireless transmission system, more especially dealing with parameters of several layers of the wireless transmission system, like the transport, network and MAC layer.

Test Bench Definition

Transmission of the video data through the wireless channels and error corruption is simulated by means of a significantly extended version of the Network Simulator (NSversion 2). In the IMEC extension of NS the protocols, which are emulated are the following: UDP and IP version 6 for the transport, network protocols and at the MAC layer, the 802.11a and the 802.11e protocols (see International Standard for Information Technology—Telecom. and Information exchange between systems—Local and Metropolitan Area Networks—Specific Requirements—Part 11: Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications, IEEE 802.11 WG, ISO/IEC 880211: 1999(E) IEEE Std. 802.11, 1999, and IEEE 802.11 WG, IEEE 802.11e/D8.0, “Draft Amendment to Standard for Telecommunications and Information exchange between systems—Part 11: Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications: Medium Access Control (MAC) Quality of Service (QoS) Enhancements”, February 2004.) where the Distributed Coordination Function (DCF) and the Hybrid Coordination Function (HCF) are simulated. For the HCF several scheduling periods are used: 8, 17 and 34 ms to cover a wide latency range, useful for interactive services like Vo IP or video games. The distance between AP and mobile device is taken as 10 m. But obviously, similar results can be obtained for other common distances in WLAN contexts.

The video codec used is a proprietary MPEG4 codec at Simple Profile (I, P frames, no B frames) optimized towards memory, (for further detail see “Motion estimation for low power video devices”, Proceedings of the IEEE International Conference on Image Processing (ICIP01), Thessaloniki, Greece, October 2001, pp. 953957. C. De Vleeschouwer and T. Nilsson, and “Costefficient Clevel design of an MPEG4 video decoder”, Proceedings of the IEEE Workshop on Power and Timing Modeling, Optimization and Simulation, pp. 233242, Goettingen, Germany, September 2000. K. Denolf, et al) where error resilience is achieved by using resynchronization video packets of fixed size and random Intra Macroblock refreshes. In the NS each video packet is encapsulated in a separate UDP packet, which in turn is encapsulated in a single MAC packet. This way, the loss of a single MAC packet causes the loss of one video packet, being all video packets independently decodable. Of course also other packeting can be investigated.

The knobs used at the application layer are the quantization parameter (QP), which determines the encoding accuracy, the synchronization video packet size and the number of forced Intra Macroblocks per frame. A lower QP value corresponds with a higher encoding accuracy and therefore both higher bit rate and quality. The use of resynchronization video packets, also called slices, allows the decoding process to resynchronize and continue the decoding process when transmission errors occur. To overcome the degradation caused by these transmission errors, the introduction of Intra Macroblocks can be enforced in order to refresh information, which cannot be predicted from previous erroneous frames.

Nevertheless, in our study we observed that the percentage of Intra Macroblocks was not a useful knob as the targeted PER are so low that the refreshment of frames by means of a small percentage of Intra Macroblocks becomes sufficient and the results are shown for 8% of Intra MB refreshment.

It is also important to note that in the current demonstration the settings for the video codec are fixed for the whole sequence, without performing at this stage an online video adaptation to the network conditions.

Of course a runtime control loop to steer these knob settings based on the channel and user data scenario's is possible also.

At the network side the main knobs we can play with are the following: modulation type, code rate, transmission power, backoff power, and Medium Access Layer (MAC) packet size among others.

In our case the selection of the video packet size coincides with the MAC packet size as each video packet is encapsulated in a separate MAC packet, so that a MAC packet loss does not affect the decoding of other independently decodable video packets, which go in other MAC packets.

There is a tradeoff in the selection of the packet size both at the application and link layers. At the application side smaller resynchronization packets allow a higher chance of resynchronization in case of errors but also introduce a higher overhead and reduce the prediction efficiency within a frame. At the MAC layer smaller packets increase the overhead (and with it the transmission energy) but they also reduce the possibility of a packet being hit by a transmission error, which relaxes the protection requirements.

To fully explore the cross layer interdependencies we run tests not only with different video coding settings (varying QP, video packet size and percentage of intra macroblock refreshments) but with several configuration of the network parameters (code rate, constellation, packet size, transmission power . . . ) Table 5 shows the range of the multimedia settings, which is taken into consideration. The range of network settings can be found in EP 1467294. The test bench used includes the Foreman sequence (medium degree of motion and bit rate requirements) and the Mobile sequence, which has a higher degree of motion and it is more demanding in terms of bit rate.

In addition the tests are performed for 8 different channel states to evaluate the impact of different transmission conditions and allow adaptability to those. These states have been shown to characterize the channel behavior well.
TABLE 5 


Application knobs 


 Sequence  Foreman, Mobile 
 QP  6, 7, 8 and 9 
 Packet size (bits)  4000, 6000, 8000, 12000 and 16000 
 Intra MB  3, 8, 20 and 25 
 refreshment (%) 
 

The NSbased simulator provides as a result for each network and video settings the energy consumed at both network source and destination, together with the maximum and average latency per video packet that the network introduces. The latency introduced at the application level together with the energy consumed at the encoding/decoding process is added to the network values.

The energy consumed at the application for the encoding and decoding process of bitstreams with varying settings is measured with a tool. The processor energy is estimated to be comparable to the energy spent on memory accesses. This estimate gives only relative accuracy (not in absolute terms) but for our main conclusions, this relative accuracy is sufficient to substantiate the concept. The maximum encoding/decoding latency per packet is measured on a Pentium IV at 2.8 Ghz, where the encoding/decoding processes are downscaled to realtime processing (30 frames per second in the case of Mobile and 25 frames per second for Foreman). For other processor platforms, the absolute values will again change, but the relative differences are expected to hold.

Video quality is measured with the Structural Similarity Index (SSIM) (see “Video Quality Assessment Based on Structural Distortion Measurement” Signal Processing: Image Communication, vol 19, No. 1 January 2004, Zhou Wang, Ligang Lu and Alan C. Bovik). We have observed that as PSNR, the SSIM value also does not correlate properly with the subjective video quality, when transmission errors occur. This is the reason why the quality axis is currently rather taken as a constraint and high video quality is imposed for all Pareto points, corresponding then to Packet Error Rates equal to zero or very low so as to keep the perceived quality high.

In an alternative setting this constraint is relaxed and latencyenergyquality tradeoffs are generated.

Simulation Results

Some benefits of the HCF scheme can be found in Appendix A as “Optimal Fixed and Scalable Energy Management for Wireless Networks”.

FIGS. 28 to 43 show certain tradeoff points obtained in the EnergyLatency axes for channel states 2 and 4, for transmission modes DCF and HCF. These tradeoff points, so called Pareto points, are extracted for each of the video coding settings (combination of QP and packet size). As mentioned previously, the quality axis is not shown as good video quality is taken as a constraint. This way, all resulting Pareto points correspond to high video quality where the Packet Error Rate (PER) is zero or very low.

The general trends, as expected, are that the higher the bit rate of the encoded sequence (corresponding to lower QPs with a higher encoding accuracy) the higher the energy invested at both network (higher amount of packets to be transmitted) and application sides. At the same time, for a constant bit rate a higher packet size implies lower energy and latency at the network side, as the overhead is lower. On the other side, when the transmission conditions get worse, big packet sizes have a higher probability of being hit by an error, introducing then more errors at the link level, which decreases in turn the energy efficiency. In our case, as the transmission distances are relatively small and the network load is low (one user) so the channel conditions are not bad enough to show this effect.

As the channel conditions get worse the more energy needs to be invested at the network side to successfully transmit the video packets with a low or zero PERs. This can be observed when comparing channel states 2 (relatively good channel) and 4 (worse channel conditions). This way, for different channel conditions the optimal points in terms of network and codec settings vary.

At the application side, lower QP values, corresponding to higher encoding accuracy) increase the encoding/decoding energy. But in terms of encoding/decoding latency, for a fix packet size, the encoding/decoding latency per packet is higher for higher QPs. This is due to the fact that the encoding/decoding process is scaled down to realtime processing and then the higher the bit rate (lower QP) the higher number of packets per frame are used and their relative encoding latency decreases.

All graphs show that there is always a large tradeoff present between energy and latency. To achieve low packet latencies a higher energy needs to be invested at the network, while if we target lower energies the global latency will increase. By playing with both multimedia parameters (QP, packet size . . . ) and network parameters (code rate, modulation, transmission power . . . ) the amount of possible tradeoffs increases, enlarging then the Energy and Latency ranges. These enlarged ranges increase in turn the flexibility of the global system, which has more possible operation points and allows higher savings in energy or latency by playing with the proper knobs. The application of DCF and HCF modes of operation creates several Pareto points for each particular video setting. This, together with the use of several application knobs (video settings) allows to extract a global Pareto curve with extended energy and latency ranges.

In addition, the channel state highly influences the resulting tradeoffs as well as the network and application settings of all Pareto points in the global Pareto curve.

The energylatency ranges that are obtained are often of factor 2 or 3, reaching even factor 4 for worse channel conditions (channel 4) and Mobile sequence while factors 1.3 to 2 are mainly obtained for Foreman.

It can be observed how the tradeoffs are larger for the Mobile sequence than for the Foreman, this is due to the larger bit rate variation among the extreme settings of Mobile (from 2 to 4 Mbps) than the one in Foreman (from 0.7 to 1.1 Mbps). Therefore, larger bit rate ranges imply larger energy and latency ranges. Similarly, the range of the energylatency tradeoffs also increases with worse channels. Both effects of increasing the bit rate or having worse channel conditions are similar in the sense of demanding higher requirements from the system.

These large tradeoff ranges can be exploited by a runtime controller that sets the appropriate knob positions based on the user QoE requirements. If battery power is low, latency can be sacrificed for the interactive services. In the other case, the user will probably prefer lower latencies. Without these crosslayer analysis and optimizations, the tradeoff ranges would not be present. Note that even the “reference” points in the graphs below already contain the crosslayer interaction between the multiuser scheduler and the WLAN baseband and frontend layers, where a runtime controller is already steering the appropriate knob positions based on channel conditions. If the latter crosslayer optimization is removed, the remaining tradeoffs will become minimal.

In the following graphs (FIG. 28 to 43) different colors correspond to different QP, while the point types correspond to different video packet sizes as follows:

QP 9: green

QP 8: cyan

QP 7: magenta

QP 6: blue

Packet size 4000 bits: ‘+’

Packet size 6000 bits: ‘x’

Packet size 8000 bits: ‘square’

Packet size 6000 bits: ‘diamond’

Packet size 6000 bits: ‘circle’

In FIGS. 40 and 42 the green curve represents the Pareto curve for the bitstream configuration selected as reference (QP 9, packet size 8000 bits) while the pink curve represents the global Pareto curve with the best tradeoffs EnergyLatency among bitstreams of different settings in terms of QP and packet size.

Crosslayer adaptation including the multimedia application layer proves to be beneficial as we can observe that for the same latency value it is possible to obtain savings of 25% in the energy, while the quality is maintained or even increased. In the same way, lower latencies can be achieved by changing the video parameters (QP, packet size). These optimal operation points are dependent on the channel state, which is being considered. Note that all these points already include the crosslayer optimization (and runtime controller) between the multiuser scheduler and the WLAN baseband/frontend (as discussed above).

Tables 6 to 9 show the energy and latency values for points in both curves corresponding to low, medium and high latencies within the given range and for channel states 2 and 4.
TABLE 6 


Reference Curve for Channel State 2. 
Reference  Wireless  Latency  Energy   Packet size 
Curve  Setting  (ms)  (Joules)  QP  (bits) 

High Latency  98  42.12  6.87  9  8000 
Medium Laten  109  37.28  7.6  9  8000 
Low Latency  NA  NA  NA  NA  NA 


TABLE 7 


Optimal Curve for Channel State 2. 

Wireless 
Latency 
Energy 

Packet size 
Optimal Curve 
Setting 
(ms) 
(Joules) 
QP 
(bits) 

High Latency 
98 
42.12 
6.87 
9 
8000 
Medium Laten 
108 
35.44 
7.19 
9 
6000 
Low Latency 
106 
27.5 
10.02 
6 
8000 


TABLE 8 


Reference Curve for Channel State 4. 
Reference 
Wireless 
Latency 
Energy 

Packet size 
Curve 
Setting 
(ms) 
(Joules) 
QP 
(bits) 

High Latency 
97 
50.77 
12.10 
9 
8000 
Medium Laten 
120 
37.28 
21.25 
9 
8000 
Low Latency 
NA 
NA 
NA 
NA 
NA 


TABLE 9 


Optimal Curve for Channel State 4. 

Wireless 
Latency 
Energy 

Packet size 
Optimal Curve 
Setting 
(ms) 
(Joules) 
QP 
(bits) 

High Latency 
49 
50.81 
9.16 
8 
6000 
Medium Laten 
107 
37.27 
15.85 
8 
6000 
Low Latency 
125 
24.69 
27.1 
6 
6000 


While the above detailed description has shown, described, and pointed out novel features of the invention as applied to various embodiments, it will be understood that various omissions, substitutions, and changes in the form and details of the device or process illustrated may be made by those skilled in the art without departing from the spirit of the invention. As one example, it will be appreciated that the test voltage driver circuitry could be separate from the array driver circuitry used to create the display. As with current sensors, separate voltage sensors could be dedicated to separate row electrodes. The scope of the invention is indicated by the appended claims rather than by the foregoing description. All changes which come within the meaning and range of equivalency of the claims are to be embraced within their scope.