CHAPTER 2 LITERATURE REVIEW 2.

1 SYSTEM MODEL In this , we motivate the problem and present the system model including assumptions regarding different aspects of the system operation. For ease of exposition, we shall use the CDMA !!! "x#$$ system as an example of a practical system suited to our proposed adaptive resource allocation mechanism, although our schemes also apply e%ually to other wireless systems that dynamically adapt wireless channel rates in response to user sending rates

Fig. ". Illustration of a cellular hop.

Fig. " depicts the wireless hop in a typical cellular networ&. It comprises a base station, mobile devices, and a buffer at the base station for each user. $he channel scheduler resides at the base station 'or in the case of CDMA !!! "x#$$, the (ase )tation Controller* and determines the rate allocated to each mobile session. For the purposes of this wor&, we focus on mobile sessions that involve $C+ bul& transfers on the downlin&. $he channel scheduler is assumed to have the ability to dynamically assign different wireless transmission rates 'termed channels* to each user in a sector. For example, CDMA !!! "x#$$ supports five different rates ranging from ,.- to "./.- 0bps per user by changing processing gain and power to trade off channel error for increased capacity.

$he main theme of this wor& is to show that $C+ throughput can be increased by exploiting this feature, in particular through careful selection of the channel rates and a mechanism to adaptively allocate different channel rates in response to $C+ state. #ate adaptation is also beneficial when multiple $C+ sessions are in progress. $hrough a detailed analysis of a single $C+ session in a multirate scenario, we show that the higher rate channel is not always utili1ed. $his immediately suggests that by virtue of statistical multiplexing gain, with a few high2rate channels and careful allocation of these channels in a multisession system, $C+ sessions can achieve nearly similar throughput as isolated sessions with dedicated channels. where it is shown that this framewor& and our proposed channel allocation schemes yield bandwidth3energy savings and higher throughput compared to a networ& with a single static channel rate.)uch a framewor& is also practically relevant. Again ta&ing CDMA !!! as an example, high2rate channels are achieved by reducing processing gain 'e%uivalently shortening the 4alsh orthogonal code*. $his ma&es them much more susceptible to interference compared to low2rate channels, and conse%uently, only a small number of users may use a high2rate channel simultaneously. For example, /! users may simultaneously use a ,.-20bps channel but only two users may concurrently be assigned a "! .520bps channel. $his is very similar to our framewor& where in several $C+ sessions contend for a few high2rate channels and hence, our results are applicable to such a system. 6ext, we present the system model and assumptions utili1ed in our wor& that capture the various aspects of a CDMA system mentioned above7 ". For simplicity, we focus on a two2rate system to explain our rate allocation framewor&. 8xtensions of the system model to three or more rates are straightforward. Denote by C! and C" the low2rate 'also called fundamental* channel rate and high2rate 'also called supplemental* channel rate, respectively. Clearly, C! 9: C". At each point in time, the scheduler decides which of the two channels, the low2rate channel 'rate C!* or the high2rate channel 'rate C"*, are to be assigned to a $C+ session. 8ach active user can always be assigned a low2rate channel; however, allocation of high2rate channels is arbitrated by the channel scheduler because there are only a few highrate channels.

. $he pac&et error probability is implicitly assumed to be a function of the assigned rate and denoted by p!'p"* when the assigned channel rate is C !'C"* . $his is an important feature representative of current wireless systems where an increase in channel rate typically comes at the cost of increased pac&et error probability. 4e assume the presence of power control to primarily combat fast fading and interference effects. $his is true in current systems where fast closed2loop power control trac&s a specified target )I6# 'or e%uivalently target F8#*. /. 4e assume no 'or a very small* buffer at the base station. <ence, $C+ experiences congestion if its sending rate exceeds the maximum assigned channel rate. 5. Finally, in the context of multiple $C+ sessions, we assume a fixed population of 6 sessions. 8ach session can be assigned a low2rate channel; however, only sessions can be assigned a high2rate channel concurrently. $his captures issues such as increased interference in high2rate channels and also allows us to study statistical multiplexing gain by varying 0. Finally, we emphasi1e that at this stage, no specific assumptions have been made regarding how the channel rates are achieved nor how they result in the specific channel error probabilities. Indeed, the specific relation is not re%uired in our $C+ model and only the actual variables are re%uired. $he channel rates and pac&et error probabilities are a function of the underlying technology that is used, e.g., adaptive modulation and spreading. 2.2 TCP-AWARE RESOURCE ALLOCATION In this , we present our channel schedulers for single2user and multiuser systems. 4e first describe the single2user scheduler, and based on this, we then propose two multiuser schedulers. 2.2.1 SINGLE-USER SCHEDULER $he $C+2aware channel scheduler we consider resides at the base station. It is %uite similar in operation to the "x#$$ scheduler described in /, with the exception that it assigns a channel rate based on the user=s $C+ sending rate. rather than its buffer content. )pecifically, whenever the $C+ sending rate exceeds 'or drops below* the current channel rate, the base station increases 'decreases* the channel rate 'accompanied by the

corresponding F8#, signal power, etc.* If the session sending rate exceeds the maximum possible channel rate, the session experiences pac&et losses due to congestion. As before, we utili1e a two2rate system as the reference example. 4e note that modern cellular systems have five or more modes, i.e., they can support up to five or more different channel rates. <owever, obtaining succinct analytical expressions for $C+ throughput even for three rates is %uite difficult. Moreover, our simulations show that there is only a small improvement in system performance by considering more than two rates. $he single2user scheduler operates as follows7 $hescheduler decides which of the two channel rates C! andC" are to be assigned based upon the user=s $C+ sendingrate. If the $C+ sending rate is below C!, the schedulerassigns a channel rate of C!, otherwise, it assigns C". Intuitively, when the $C+ sending rate is small, a lower F8# is essential 'to avoid time2outs, etc.*. $he scheduler can exploit the &nowledge of the low $C+ sending rate to trade off channel rate for channel integrity. At higher $C+ sending rates, it is more appropriate to assign a larger channel rate at the expense of a higher F8#. $his is because, even though pac&et loss probability due to channel errors increases, a larger channel rate prevents pac&et loss due to congestion, which would have happened with probability one were the rate not increased, allowing $C+ to transmit at high rates for a longer time. 4e note that there are several other features of $C+, for example, time2out values, window si1e, sending state, etc., which could potentially be utili1ed to improve upon our proposed scheduler. Incorporation of such features however, would ma&e the scheduler complex to implement as well as to study. >ur aim here is to propose a system that involves minimal modifications to schedulers used in current technologies li&e CDMA !!! "x#$$ and 8?2D>, and can be studied analytically. From this perspective, we believe that our proposal to incorporate &nowledge of only $C+ sending rate satisfies both goals. 2.2.2 MULTIUSER SCHEDULER In this , we propose two flow2level channel schedulers, which are extensions of the single2user scheduler described in the previous , for channel allocation among multiple $C+ sessions. 4e assume that all users have the same propagation delay. $his is not unreasonable since all ma@or cellular networ&s deploy split2 $C+ on the downlin& implying that $C+ sessions on the downlin& originate at a common proxy and, hence, share a common path. A

naive extension of our single2user scheduler would be to have a dedicated high2rate channel for every user in the system. <owever, our analytical and simulation results for single2user system show that the high2rate channel is not fully utili1ed implying that such an extension would be wasteful of system resources, i.e., high2rate channels. <ence, we assume that there are only a few high2rate channels as is the case in CDMA !!! systems and propose flow2 level scheduling strategies to maximi1e $C+ throughput in a multiuser scenario. 2.2.2.1 A -EMPT SCHEDULER $he A28M+$ , which has a control parameter A wor&s as follows7 ". Bpon commencement, a $C+ session i is assigned a fundamental channel with rate C!. 6ote that all 6 sessions can be assigned fundamental channels. . )imilar to the single2session case, the scheduler must ma&e a decision on allocation of a high2rate channel whenever a $C+ session=s sending rate reaches C!. /. In such a scenario, if a high2rate channel is free then it is allocated to the re%uesting session. >therwise, one of the 0 users occupying the high2rate channels is preempted with probability C and the released channel is allocated to the re%uesting $C+ session. 6ote that in this case, the preempted session faces congestion loss with probability one and halves its sending rate. Alternatively, the re%uesting session is denied the high2rate channel, in which case it is the re%uesting session that experiences congestion and drops its sending rate by half. 5.4henever a user in a high2rate channel drops its sending rate 'triggered by either a congestion loss or a random channel error*, the user releases the highrate channel and gets assigned a fundamental channel. 4e need to specify the preemption probability A in order to completely specify the scheduler. Furthermore, it is clear that A, the number of supplementary channels as well as the channel rates will all impact $C+ throughput. 2.2.2.2 -AWARE Schedule $he D2A4A#8 scheduler is similar to the A28M+$ scheduler, with the difference that instead of randomly selecting a user occupying the supplemental channel for preemption, it

chooses a user with the largest $C+ window si1e among those whose rate is above a factor D of the pea& capacity C". $he intuition behind the D2A4A#8 scheduler can be explained as follows7 Eet us assume that user @ has a high2rate channel and F@'t* is the sending rate of user @ at time t t'C! 9 F@'t* 9 C"*. 4e also assume that at time t, user i re%uests a high2rate channel, i.e., Fi't* :C!. Consider the following scenarios7 ". If F@'t* is close to C!, then there is no throughput gain in preempting user @ to allocate the high2rate channel to user i. In fact, the overall systemthroughput may even degrade if user @ is preempted. . If F@'t* is very close to C", then after a short period of time, user @ will reach C" and face congestion with probability one. 4hen that happens, the high2rate channel will stay idle until another user ma&es a re%uest. In this case, system throughput can be improved by preempting user @ to accommodate user i at high rate. D2A4A#8 tries to distinguish between these scenarios by predicting how much longer a high2rate user can send at high rate before reaching C". D2A4A#8 uses a threshold D!'C!3C" 9: D 9: "* to decide when to preempt a high2rate user @ by considering how close F@'t* is to C". $he scheduler operates similar to A28M+$ except when contention occurs.Assume user i re%uests a high2rate channel when all highrate channels are occupied. Eet @G denote the user that has the highest sending rate, i.e., @G : arg max"9:@9:6 F@'t*. $hen,D2A4A#8 wor&s as follows7 ". If F@G't* 9: D C", then user i=s re%uest for a high2rate channel is denied. . >therwise, user @G is preempted and the released high2rate channel is assigned to user i. $herefore, we have a multiuser scheduler that ta&es into consideration the sending rate of both low2rate users and high2rate users in ma&ing scheduling decisions. Furthermore, note that D: " corresponds to A :! in the A28M+$ scheduler since in both cases this would result in no preemption. )imilarly, D: C!3C" corresponds to A :". by varying the parameters

D and A, we study a wide range of behaviors when either ignoring or incorporating $C+ rate information into channel allocation decisions. It is worth mentioning that since we consider longlasting $C+ flows in a homogeneous environment where all sessions experience the same propagation delay, the above schedulers are inherently fair. All sessions are statistically treated the same by the schedulers, and hence, they achieve the same steady2state throughput. 2.! SINGLE TCP SESSION ANALYSIS In the previous , we presented a $C+2aware channel scheduler for single and multiple sessions. 6ext, we present an analytical model for the single2session case that explicitly captures the interplay between $C+ and the scheduler in a two2rate system. )uch a model can e useful to %uic&ly explore system behavior instead of utili1ing lengthy simulations. From a $C+ flow perspective, such a system results in two distinct operating regimes with different channel capacities, #$$s and pac&et error probabilities. In this , we present a $C+ model for such a system in the single2session case that enables us to study the impact of dynamic channel allocation on $C+ throughput. 8xtending the model to multisession case is an ongoing wor&. 2.!.1 TCP M"del 8xisting $C+ models typically assume that the #$$ and pac&et loss statistics are independent of $C+ dynamics in throughput calculations. <owever, in the wireless environment considered in

Fig. . $C+ window si1e evolution over a variable rate channel.

this wor&, there exists a strong correlation between the scheduler and $C+ sending rate. In particular, the channel capacity C that affects #$$, as well as pac&et loss probability p are functions of the $C+ sending rate. As an illustration, Fig. depicts the evolution of window si1e of $C+ in steady state when serviced by the proposed $C+2aware channel scheduler. $he scheduler assigns rates based on the $C+ sending rate and as can be seen, this in turn affects the window growth rates.. for a single fixed rate channel as a starting point and develop a model that accounts for the two2rate regime. 4e present a more detailed explanation of our model below7 ". 4e assume that the $C+ version is $C+ #eno and model the $C+ window growth in steady state as a fluid process where the window si1e grows linearly in the absence of loss. . $he sender is assumed to always have data to send and, for analytical tractability, we ignore time2outs and slow start. /. Eet 4't* denote the window si1e of $C+ at time t and #'t* the #$$ at time t. In the absence of a buffer, if the scheduler is in mode i : !; " at time t, we approximate the #$$ #'t* as #'t* : #i Ha * E3Ci; where a is the round2trip propagation delay, E is the pac&et length, and Ci is the channel capacity in mode i. 5. Eet F't* denote the instantaneous $C+ sending rate at time t in bits per second. $hen, F't* : 4't* 3#'t* . .. During congestion avoidance, the $C+ window si1e increases by roughly one pac&et 'E bits* every #i seconds in mode i when there is no pac&et loss. 4e approximate this in our fluid model with a linear growth rate of E:#i. Conse%uently, the sending rate grows at a linear rate of E:# i -. As in and I", we assume that the channel losses can be modeled by an

inhomogeneous +oisson process with rate piF't*, i : !; " at time t. J. If the $C+ sender is not constrained by the receiver window, then $C+ experiences congestion with probability " when its sending rate, F't*, exceeds C". <owever, if $C+ is constrained by the receiver window si1e, the sending rate stops increasing once it reaches the receiver advertised window and it experiences only channel2related losses. 4e need to

approach the problem differently depending on whether the two modes 'channel rates* satisfy either C" 9: C! or C" K C!. For the important case- of C" 9: C!, we derive expressions for mean $C+ throughput for both the unconstrained rate case as well as the constrained rate case. 4e have also obtained expressions for the $C+ throughput for the case when C" K C! but omit them here due to lac& of space as well as the fact that most performance improvements are obtained in the region C"9: C!. 2.# NUMERICAL RESULTS $his is devoted to %uantifying the throughput gain obtainable with $C+2aware rate adaptation, verifying the accuracy of the analytical model as well as exploring the impact of the two resource contention mechanisms on multiple $C+ sessions in a CDMA system. (efore evaluating the performance of the rate adaptation scheme, we must address the issue of how rate adaptation by the channel scheduler affects both channel capacity and pac&et error probability. 6ote that our framewor& does not assume any specific dependencies between the various channel rates and pac&et error probabilities, rather it assumes these are predetermined input variables. #ate adaptation in CDMA systems can be achieved by changing any of three variables7 error coding rate, the spreading factor, or modulation scheme. In this wor&, we study the impact of the first two variables, though our framewor& is e%ually applicable to any other means of controlling the data rate. 4e initially assume that the channel rate and pac&et error are controlled solely by changing the coding rate. $his is primarily to provide a simple basis to evaluate the accuracy of the model as well as showcase its utility in a single2user framewor&. $his is representative of current CDMA networ&s where spreading factor is the dominant control &nob to adapt data rates. 4e evaluate the gain in $C+ throughput and resultant trade2off with energy consumption as a function of the spreading factor., we evaluate a system with multiple $C+ sessions, in particular number of channels, etc., on $C+ throughput. In passing, we note that another potential control factor in wireless networ&s is the retransmission mechanism typically deployed at radio lin& layer to mitigate high F8#s by retransmitting erroneous frames.

>ur model can easily incorporate the impact of lin&2layer retransmissions as a trade2 off between channel rate and pac&et error probability similar to the coding rate studied here, though, it cannot account for the latency introduced by lin&2layer retransmissions. <owever, extensive measurements conducted over a commercial "x#$$ networ& by Mattar et al. found that the impact of rate changes of the scheduler on $C+ sending rate is far more dominant, while there is little or no correlation between $C+ #$$s and the lin&2layer retransmissions primarily due to large #$$s and very fast retransmissions. <ence, we do not study the impact of lin&2layer retransmissions in this wor&. 2.#.1 PAC$ET ERROR PRO%A%ILITY& THE VARIA%LE CODING CASE As mentioned previously, we assume that the capacity Ci and pac&et error probability pi in mode i are functions of the coding rate Li. <ence, for a given bit error probability, the $C+ throughput is a function of the two coding rates, i.e., F'L!,L"*. $he relation between the capacity and coding rate is straightforward and is given by Ci :Li . CG, where CG is the un coded channel capacity. $he pac&et error probability, however, is strongly dependent on not @ust the coding rate but also the coding scheme used. Conse%uently, one must either choose a specific coding scheme, or resort to bounds on the achievable pac&et error probability for a given coding rate. >ne such bound is the Milbert2?arshamov bound. $his was used in and we also use it as an approximation. $he Milbert2?arshamov bound is a bound on the parameters of a code of length ( and information bit length 0. It specifies that there exists a minimum <amming distance d between any two codewords 2.#.2 ADAPTIVE CODING EVALUATION >ur evaluation of the $C+ model by comparing its accuracy against simulations of the $C+2aware channel scheduler implemented in ns2 . $o be concrete in this, we set the raw channel rate CG to " N 0bps and the two2way propagation delay a to !! ms. $C+ pac&et si1e isset to ",! 5 bits. A $C+ pac&et is divided into radio bloc&s of si1e .- bits for transmission over the wireless channel. $he same parameters were also used for the ns2 simulation. $C+ #eno was chosen 'since the model fits that version best* and pac&et errors were assumed to be independent and identically distributed.

)ince we do not account for time2outs in our model, we only simulated scenarios with large window si1e that result in few time2outs. Indeed, a previous study has recommended the use of large window si1es in cellular networ&s to precisely avoid such time2outs due to bandwidth oscillations. $he duration of each simulation run was /!! s. $he bit error probability was held constant for the duration of the simulation, which is true under perfect power control. <ence, the pac&et error probability is solely a function of the coding rate. 4e ran ! simulations with different random seeds for each data point and the results reported are in the ,. percent confidence interval. It is worth mentioning that we limited ourselves to low pac&et error probabilities in all the scenarios. $he reasons for this are twofold. First, $C+ is &nown to perform well only for low pac&et errors 'less than . percent*, and hence, it represents the region of interest. $he second reason has to do with modeling the pac&et loss process as an inhomogeneous +oisson process which is reasonable only for low pac&et error probabilities. In all the scenarios considered in this pro@ect, the model matches the simulation results closely with errors typically less than . percent 'please see for error plots*. $o demonstrate the importance of capturing the correlation between $C+ and the scheduler as well as the presence of two states, we plot in Fig. 5 the difference between the two2rate model 'which we have shown above to be accurate* and a single2rate model that ta&es only one set of parameters due to coding rateL! into consideration 'please see for the result of considering L"*. 4e observe that there exist regions where there is substantial difference in predicted throughput ' ! to /! percent*. <ence, a singlerate model may not be always feasible to tune performance in such systems.

2.#.! OPTIMAL CODING RATES

Fig. 5. Comparison between two2rate and single2rate models with parameters 'C!, p!, #!* for bit error probability 'pe* : "!2 . 2.#.!.1 TWO-RATE SCHEDULER Assume the adaptive coding case, channel rates are controlled through the coding rate. $herefore, the scheduler is essentially switching between two coding rates L! and L". Clearly, the particular choice of L! and L" affects the achieved throughput. Intuitively, if pac&et error probability is close to !, then increasing the coding rate increases throughput as well. $he reason is that, in this situation, channel rate increases as a linear function of the coding rate, whereas the pac&et error probability remains negligibleresulting in increased throughput. <owever, if the pac&et error probability is large, then increasing the coding rate decreases throughput and eventually reduces it to !. Conse%uently, one expects the existence of a coding rate that maximi1es $C+ throughput. )uch behavior has been previously observed for the single2rate case in for finite capacity and and for infinite capacity models.

4e extend these previous results by identifying the optimal pair of coding rates within the framewor& of an adaptive $C+2aware channel scheduler.

$A(E8 "7$wo2#ate )cheduler7 >ptimal AdaptiveCoding #ates 'CG: " N 0bps*

$A(E8 7$wo2#ate )cheduler7 Main of Adaptive Coding $able " presents the pair of optimal coding rates OC! and OC" obtained from both the model and simulations that maximi1e the $C+ throughput F'Co; C"* and the corresponding throughput for different target bit error probabilities. >bserve that in most cases, there is a close match between the optimal coding rates 'and corresponding throughput* obtained from the model and those from simulations. In order to %uantify the benefits of using two coding rates, $able presents the relative increase in throughput as a conse%uence of using adaptive coding compared to a singlecoding rate for each bit error probability. <ence, for each bit error probability, the gain factor was computed by comparing throughput at the optimal

coding rates for adaptive and static coding, respectively. In all the cases, we see that the optimal throughput with adaptive coding is at least e%ual to that of static coding and higher in several situations, with improvements of up to J percent. Intuitively, the reason adaptive coding yields a gain in throughput over static coding even with the same target bit error probability is that the channel scheduler with adaptive coding exploits &nowledge of $C+ sending rate. )pecifically, when $C+ has a small window, it sends at a low rate.<ence, the channel scheduler can offer a smaller rate to the source but with a lower error probability. As the rate increases, the scheduler switches to a higher rate to cope with $C+. $his is not possible in the single2rate case. 2.#.!.2 THREE-RATE SCHEDULER throughput gain is possible when the wireless scheduler can transition between three channel rates. In our current evaluation framewor&, this essentially means that the scheduler is switching between three coding rates L!, L ", and L . $he three2rate $C+2aware scheduler with channel rates 'C!; C"; C * that correspond to coding rates 'L !; L "; L *. It is worth mentioning that the ns2 simulations to identify the optimal rate configuration of the adaptive coding scheme is considerably time consuming even for the simpler two2rate scheduler. Computing $C+ throughput via ns2 for each combination of rates ta&es almost a day for ! runs of simulations. $herefore, we have only simulated a three2rate scheduler in this pro@ect. <owever, we expect to see similar results for channel schedulers with more than three channel rates.

$A(E8 /7$hree2#ate )cheduler7 >ptimal AdaptiveCoding #ates 'CG: " N 0bps*

$able / summari1es simulation results for three2rate scheduler for different (8#s. In the table, 'L !; L "; L * denotesthe optimal coding rates corresponding to three channel rates used by the scheduler that maximi1e $C+ throughput. 4e have also computed the gain factor representing the percentage of throughput improvement compared to static coding throughput as defined before. An interesting observation is that, for all simulated cases, the throughput gain of three2rate scheduler is roughly e%ual to that of the two2rate scheduler considering the margin of simulation error. In other words, most of the throughput improvement is already achieved by using only two rates and increasing the number of channel rates does not further improve the throughput. 2.#.# OPTIMAL PROCESSING GAIN Assumed that the trade2off between the channel capacity and F8# was controlled by the coding rate. In CDMA networ&s, the dominant #F control variable that governs this trade2off is the spreading factor. $he spreading factor is defined as the ratio of the CDMA chip rate to the actual data rate and relates the channel capacity and the bit error probability in the following manner. Eet 4 denote the chip rate of the CDMA system and ri the spreading factor allocated in mode i. $ypically, the bit error probability is extremely sensitive to )I6# and hence, in practice, some amount of power control is re%uired to prevent a steep rise in the error probability when the spreading factor is decreased. )pecifically, the chip energy is boosted by a small factor that is a function of the spreading rate to prevent a sharp drop in the %uality of service.

$A(E8 57$wo2#ate )cheduler7)preading Factors 'CG : "./7- 0bps*8nergy +rofile 8"

$A(E8 . 7$wo2#ate )cheduler7 Max #ate Configuration

single2 and the two2rate cases. $he gain in $C+ throughput and difference in energy consumption between the adaptive and static cases, under their respective optimal configurations, are shown in $able . for both energy profiles. $here are a number of interesting observations regarding the results. From $able ., the analytical model indicates that a small increase in energy consumption, on the order of / to 5 percent, in con@unction with a two2rate strategy boosts $C+ throughput by "! to ". percent as compared to the single2rate scenario even though the allowed pea& channel capacities are the same. $he rate gain and energy saving predictions of the simulation are comparable to those of the model for 8" and 8 . Interestingly, simulation predicts higher benefits than the model for energy savings in the case of 8 . In fact, it predicts a positive energy savings of 5. percent.$he differences in energy prediction for 8 are due to the model and simulation choosing different optimal configurations in the single2rate case, while yielding the same configuration in the two2rate case. )pecifically, the model overestimated 'by about , percent* the $C+ throughput in the single2rate case for the highest rate configuration of "./.- 0bps, causing it

to pic& that as the optimal. In comparison, the simulation chose the "! .520bps rate which has higher normali1ed energy consumption. 2.#.#.1 THREE-RATE SCHEDULER 4e have also simulated a three2rate scheduler that utili1es all three rates PJ-.N, "! .5, "./.-Q 0bps and the same energy profiles used in the previous . $able - summari1es the results for throughput gain using the three2rate scheduler. Comparing the throughput and energy gains for the two2rate system in $able ., we see that they are almost the same, indicating that two channel rates yield a large fraction of the benefits. 2.#.' MULTIUSER SCHEDULER PER(ORMANCE In previous , we showed that by exploiting dynamic rate adaptation, it is possible to improve system throughput by "! percent to ! percent for a single $C+ session. 4e next explore the impact of rate adaptation when multiple $C+ sessions are present by studying the performance of the two proposed high2rate channel allocation algorithms7 A28M+$ and D2A4A#8 . 2.#.'.1 )-EMPT PER(ORMANCE EVALUATION

$A(E8 -7$hree2#ate )cheduler7 )preading Factors 'CG: "./7- 0bps* For the purposes of our evaluation, we utili1e the same scenario as in .5.5 for a

CDMA system. )pecifically, we assume that rate adaptation is carried out by varying processing gain. Channel error probabilities were determined using the energy profile 8" to be '!, J.J, N.J, 5/. *G "!25 for channel rates of P,.-, J-.N, "! .5, ./.-Q 0bps, respectively. Fig. . plots the $C+ throughput as a function of A and the number of supplemental channels 0 for 6 : "! users. $he supplemental channel rate was set to "./.- 0bps in Figs. .a, .b, and

.c, and "! .5 0bps in Fig. .d. $he figures show the evolution of throughput for fundamental channel rates of "! .5, J-.N, and ,.- 0bps. $hese three scenarios provide several interesting observations regarding the role of A, the channel rates C!, C", and the number of supplemental channels 0, which help explain the behavior of the A28M+$ scheduler7 ". For a fixed C, the $C+ throughput is obviously an increasing function of 0. <owever, the throughput gain diminishes with increasing 0. $his can be directly attributed to the statistical multiplexing gain. Application of the single2session analytical model developed in . reveals that the supplemental channel utili1ation is around 5. and J! percent when the fundamental channel rates are "! .5 and J-.N 0bps, respectively. $his indicates that the high2rate channel is not continuously occupied, especially in the first case and, hence, can be shared to a certain extent among multiple sessions. Conse%uently, initial increments of supplemental channels 'from 0 : "* yield almost proportional increments in throughput because it reduces contention. <owever, throughput saturation sets in once 0 is large enough to satisfy the high2rate utili1ation re%uirements7 the smaller the single2session high2rate utili1ation, the faster the saturation, and hence, fewer channels are re%uired. For example, when the fundamental channel rate is "! .5 0bps 'Fig. .a*, almost all throughput gain is obtained with about five supplemental channels since $C+ sessions utili1e it at most 5. percent of the time. 4e shall explore this further in -.../. . For small values of 0, the throughput is sensitive to the choice of C and its behavior depends on the fundamental and supplemental channel rate combination. As 0 increases, the throughput becomes less sensitive to C since supplemental channels are increasingly available when re%uested and the contention li&elihood decreases. It is worthwhile to further explore the detailed behavior of $C+ throughput as a function of A for small values of 0.For ease of exposition, consider 0 : " in a low2loss system,

Fig. .. Impact of A on $C+ throughput for 6 :"!. 'a* C! : "! 75 0bps and C" :"./7- 0bps. 'b* C! : J-.N
0bps and C" :"./.- 0bps.'c* C! : ,.- 0bps and C" :"./.- 0bps. 'd* C! : J-.N 0bps and C" R "! 75 0bps.

<owever, this can result in lower efficiency compared to A: ! if the supplemental channel rate is large, since the session may get preempted before it can sufficiently increase its transmission rate 'see Fig. -*. $o summari1e, the trade2off between these two factors, which in turn is determined by the relative values of C! and C", governs the $C+ throughput as a function of A.$his behavior is exemplified in Fig. .. 4hen C! 99C" 'e.g., I,.-, "./. 0bps in Fig. .c*, the high2rate channel utili1ation dominates since a session in the high2rate channel can gain significant increase in throughput compared to the fundamental channel. Furthermore, since the fundamental channel is small, re%uest rate for the high2rate channel is large enough to avoid any potential loss of occupancy. Conse%uently, A:! is optimal because it allows a session to fully utili1e the channel. As C! increases relative to C", the gain in throughput from utili1ation of the high2rate channel diminishes 'since C! is already large* and re%uest rates reduce 'because sessions spend more time in C! as described by our model earlier*. <ence, the high2rate channel occupancy factor becomes more important. Figs. .a and .d illustrate this scenario. <ere, C! K C", and A : " maximi1es occupancy, which yields the optimal throughput. An intermediate case is C! : C" 'Fig. .b*, where the two

factors balance each other. $his results in a $C+ throughput insensitive to A. 6ext, we loo& at the impact of A on $C+ throughput gain. In Fig. J, we have plotted the maximum throughput gain of A28M+$ across different channel configurations for 6 :"! users and various supplemental channels 0.

Fig. -. Impact of A on supplemental channel utili1ation.

Fig. J. Impact of A across all scenarios.

>ne can clearly see that the A28M+$ scheduler, and in particular the value of Ais influential when the gap between C! and C" is very large 'C! : ,.- and C" : "./.- 0bps*, resulting in throughput gains of -! to ,! percent by appropriately tuning A. <owever, when C! is closer to C" 'AC": *, performance is almost insensitive to A with small gains of about 2. percent.

2.#.'.2 *-AWARE PER(ORMANCE EVALUATION Miven the relative insensitivity of $C+ throughput to the A28M+$ scheduler when the gap between C! and C" is not large, we next turn to the more sophisticated scheduler D2A4A#8 . 4e studied the impact of the D2A4A#8 scheduler on $C+ throughput as a function of D and found it to have a similar dependency as the A28M+$ scheduler on the gap between C! and C" as well as number of supplemental channels. $his results are shown in Fig. N. Fig. , presents a comparison in performance between D2A4A#8 and A28M+$ schedulers across different rate configurations, number of users, and supplemental channels 'the numbers in brac&ets represent '6,0**. It shows the gain achieved in throughput with the best configuration of the D2A4A#8 scheduler through appropriate selection of D. when compared against the best configuration of the A28M+$ scheduler. )omewhat surprisingly, we observe that in all scenarios the D2A4A#8 scheduler yields little benefit over the A2 8M+$ scheduler, with about 2. percent improvement. $his leads us to con@ecture that $C+ throughput in such situations is relatively insensitive to channel allocation strategy when C! and C" do not differ by a large value and most of the gain in throughput is achieved through rate adaptation. 2.#.'.! MULTIPLE+ING GAIN WITH MULTIPLE (LOWS $his leads us to the final aspect of the study of benefits of rate adaptation in $C+ multisession systems, specifically resource efficiency. As was alluded to earlier when discussing the A28M+$ scheduler, results from the single session analytical model indicate that in a multirate system, the high2rate channel is not fully utili1ed. $his can lead to significant savings. In CDMA systems, high2rate channels are achieved by reducing the orthogonal 4alsh code length. )ince they are highly susceptible to interference, it limits the number of simultaneous high2rate channels that can be supported 'the higher the rate, the fewer the channels*, which in turn affects capacity. Conse%uently, it is desirable to provide throughput gains while at the same time efficiently utili1ing high2rate channels.(y exploiting the statistical multiplexing gain inherent in rate adaptation, we demonstrate that one can achieve similar or higher

throughput as with single2rate channels but with a lower total bandwidth. Fig. "! presents the throughput and bandwidth savings when using multirate channels in a multiuser system

Fig. N. Impact of D on $C+ throughput for 6 : "!. 'a* C! : J-.N 0bpsand C": "./.- 0bps. 'b* C! : ,.- 0bps and C" : "./.- 0bps.

Fig. ,. $hroughput gain and energy saving as function of 0. 'a* +rofile8" 7 C : "./.-, 'C! : "! .5,C" :"./.-* 0bps. 'b* +rofile 8 7 C : "! .5, 'C! R J-.N, C" : "./.-* 0bps.