Rotating cluster mechanism for coordinated heterogeneous MIMO cellular networks

2.1 Layout 1: six-cell layout (hexagonal-shaped cooperating area)

In our first HetNet model, the coverage area of the six cells per site overall forms a hexagonal-shaped region. Each macro BS covers a 60° angle of the area with a directional antenna. The macro site is surrounded by 12 low-powered pico BSs that form picocells overlaying the macro coverage area (see Fig. 1). Of the six cells per site, two adjacent ones are coordinated at any given time to form an effectively larger cell area. The picocells also coordinate within whatever cluster that the macrocell they overlay is part of. Without loss of generality, we may consider any arbitrary macro site (with coverage area shown in green) and the clusters it participates in (shown by the red dashed lines). Therefore, the BSs of any macro site contribute to three different clusters.

As depicted in Fig. 1, two different patterns of clustering are possible, in which different adjacent cells cooperate with each other. All cells within each thick red dashed hexagon coordinate signals from their BSs to form a cluster; one example cluster in each pattern is emphasized in the figure for clarity. As Fig. 1a depicts, those users in a cluster that are located near the border of the cluster (for example, at location “1”) experience the poorest channel conditions from the BSs in the cooperating set. By rotating the clustering pattern by 60° around any macro site (see Fig. 1b), those previously poor-coverage users are now in the middle of the cluster (i.e., they will have better channel gains or higher achievable rates). Therefore, most users will have the opportunity to have a higher chance of being scheduled and to achieve reasonably good data rates for a fraction of the overall transmission time. With users being scheduled primarily during their most favorable clustering pattern, their corresponding rates will be higher than otherwise. Averaging the throughput over all transmission periods and clustering patterns, the overall achievable transmission sum rate of the users will be improved.

There are K users uniformly distributed over the coverage area of each macro site, each user equipped with M receive antennas. K_c(i) is the number of users assigned to cluster c(i), from which U_c(i) users are served, where i refers to the ith pattern of clustering. Each cluster transmits coordinated data signals from all its BSs to its scheduled users.

2.2 Layout 2: three-cell layout (clover-leaf-shaped cooperating area)

For the second HetNet model, which is more commonly used in LTE-advanced design [32] and is called a clover-leaf model, each cell in a macro site is covered by a high-powered BS, which is located at a corner of the cell. The directional antenna at a macro BS covers a hexagonal-shaped cell within the angle of 120°. Each macrocell is overlaid by four low-powered omnidirectional pico BSs. These are located near the four edges of the macrocell that are the most distant from the macro BS, as depicted in Fig. 2. Any three adjacent macrocells and their constituent picocells may form a cluster, if the macrocells share a corner that is not a site. Therefore, considering an arbitrary macro site and its corresponding three macrocells (shown in green in Fig. 2), the macro BSs may belong to two or three independent clusters (shown by the red dashed lines).

As depicted in Fig. 2, five different patterns of clustering are possible. We again highlight one example cluster in each pattern for clarity. Those users that are located near the edge of the cluster experience poor channel conditions from the BSs in cooperating set. Consequently, their achievable rates will be smaller compared to the users in the middle of the cluster. By rotating the clustering pattern (see Fig. 2b), a portion of those previously poor-coverage users are now in the middle of the cluster, and some of the users, previously located at the middle of cluster, are now near the edge of the cluster. To put the remainder of the cluster-edge users in Fig. 2a near the middle of a cluster, more rotations are required, which are depicted in Fig. 2c–e. Therefore, in this layout after five intervals of rotation all users have the opportunity of being at least once in the middle of the cooperating area; they thus have a higher chance of being scheduled and achieving reasonably good data rates.

There are K users, each equipped with M receive antennas, uniformly distributed over each cell. This is in contrast to the six-cell layout, which has K users distributed over the coverage area of the macro site. Thus, for the three-cell layout, there are 3K users per site, i.e., K_c(i)=3K.

For both layouts, other patterns of clustering could in theory be used, e.g., by coordinating more macrocells together in a cluster. However, please consider the corners of each macrocell that do not contain a macro BS site. These locations have the worst SINR when no coordination occurs. The patterns that we use cluster the smallest possible number of macro BSs such that it allows each of those corners to be in the center of the cluster in one of the patterns. At the same time, the duration between any given corner being in the center (as the scheme rotates through the patterns) is also the smallest possible for the number of macrocells per cluster being used. (Note there are two of these corners per macrocell in the six-cell layout, and five such corners in the three-cell layout, hence leading to two and five patterns, respectively, for the layouts.)

2.3 Cluster rotation in general network layouts

There are, in general, two “rotation” aspects to cluster rotation. The first can be viewed as a physical rotation. Please note the highlighted cluster in Figs. 1 and 2 (denoted by dashed and solid vertical lines, respectively). As the cluster patterns change, that cluster, in a sense, can be imagined as rotating around some location in the network. In the two cases depicted by Figs. 1 and 2, that location is the macro site in the middle of each subfigure, though this need not be the case in general. The second aspect of rotation is the periodic rotation through a set of cluster patterns, in a round-robin fashion. This latter aspect is more general to any arbitrary cell layout. The first aspect may not necessarily be applicable, or at least quite so readily visible, as the second. For example, the five patterns in Fig. 2 could be ordered arbitrarily. If so, the physical rotation aspect would not be as apparent, but the rotation through the set of (re-ordered) patterns would still occur.

While we investigate two regular grid-like cell layouts herein, the concept of cluster rotation can also be applied to more general irregular layouts. For such irregular layouts, it would first be necessary to determine sets of BSs in the network for coordination and then assign different clustering patterns to them. This may not be as simple as with a regular cell layout, but remains feasible, given a set of BS locations and coverage areas and/or where interference results without coordination. Voronoi diagrams of order n [36] could be of use to locate regions of coordinated BSs, by identifying the n nearest BSs at any given location; the distances should also be weighted based on the type/tier of each transmitting node. The system can then rotate through those patterns just as in this work.

2.4 Complexity comparison of dynamic clustering and rotating clustering

Fully dynamic clustering (whether the scheme in [30] or otherwise) results in significantly higher overhead in computational load and signaling. The system must determine and exchange possible choices of BSs for each user, run some sort of optimization or other routine to determine the choice of which BSs to serve which user, and finally communicate these choices across the network and to the users. This could occur potentially as often as every scheduling interval, though the system could also perform these operations less frequently. In comparison, almost none of those computations are required with rotating clustering since the sets of clusters are predetermined beforehand and known at all transmitting nodes. The additional overhead beyond that of static clustering is simply the same as the last stage of fully dynamic clustering, i.e., to periodically inform the users what their new cluster will be.

Furthermore, there are additional savings in complexity in regard to cell association. With rotating clustering, the association of a user to a specific anchor BS has much less impact on the network’s operation (disregarding the context of high user mobility and/or handoff, which are outside the scope of this work). Note that a user receives data from a macrocell and all picocells overlaying that macrocell. Borders between macrocells (where the received power from the BSs of those cells are equal) are statistically identical; at times, that cell border may also be a cluster border, while at other times, it will not. Thus, a complicated cell association scheme is not required. Whether a user chooses an anchor BS by closest distance, highest average received power, adding on a tier-dependent association bias factor, etc., the performance of the scheme is by and large unchanged. Essentially, users can be considered more to be associated with a cluster rather than with an individual cell; in terms of performance, it is largely equivalent to associate with any one of the cells in the cluster. There may still be, for example, considerations of offloading traffic, but these would now be between clusters rather than between cells. In any event, such factors are beyond the scope of this paper.

2.5 Achievable weighted sum rate maximization and user scheduling

z_c(i),k(b,m,n) represents small-scale frequency-flat Rayleigh fading with an i.i.d. complex Gaussian random variable distributed as \(~\mathcal {CN}(0,1)\). R _m is the reference distance², and Γ₀ is a scaling factor controlling the reference signal-to-noise ratio (SNR) at a distance of R _m in the boresight direction of the directional antenna. The distance between user k and BS b in cluster c(i) is represented by d_c(i),k(b), and α(b) is the path loss exponent for BS b. P _t (b) is the transmit power of BS b, and ρ_c,k(b) denotes the log-normal shadow fading coefficient with standard deviation σ _ρ . The antenna pattern A(θ,b) of a macro BS, where θ is the angle between the direction of interest and the boresight of the antenna at BS b, is defined as described in [32, 37]; A(θ,b) is equal to unity for pico BSs with omnidirectional antennas.

\(\mathbf {Q}_{c(i),\pi _{c(i),k}^{j}}\) is the transmit covariance matrix for user \(\pi _{c(i),k}^{j}\) in cluster c(i), and \(\mathbf {V}_{k-1}^{0}\) are orthonormal basis vectors for the joint null space of H_k−1 for the users before \(\pi _{c(i),k}^{j}\) in the encoding order; \(\mathbf {V}_{0}^{0} \triangleq \mathbf {I}_{B_{c(i)}N}\).

The throughput maximization criterion results in the selection of a scheduled vector of users that achieves the largest sum rate among all possible vectors of users. Those users who have better channel gains have a higher likelihood to be selected by an MT scheduler. Thus, users with poorer channel gains may be very infrequently (and potentially never) selected by the scheduler, which is not fair. In PF scheduling, each user has a weight related to its priority for being chosen by the scheduler. The scheduler adjusts each weight based on the average achievable rates in the user’s history. A PF scheduler chooses those users whose instantaneous rates relative to their average rates are better than the others and uses a weighted sum rate as its scheduling metric, i.e., the combination of those users with maximum weighted sum rate will be chosen to be scheduled. If a user has been selected by the scheduler often, its weight for the next interval will be decreased (as its average rate increases), i.e., its chance to be chosen in the next scheduling interval diminishes. Meanwhile, another user with a worse channel matrix may have more opportunity to be scheduled in the next interval simply by having higher weight. Using this method provides more fairness in the network among all users.

where w_c(i) is the number of macrocells in cluster c(i).

To solve the optimization problem in (6) using (5) as \(R_{c(i),\pi _{c(i),k}^{j}}\), we must consider the \(\mu _{c(i),\pi _{c(i),k}^{j}}(t)\) weights when calculating \(\mathbf {Q}_{c(i),\pi _{c(i),k}^{j}}\) using the water-filling algorithm, which allocates power over the eigenmodes of the block-diagonal matrix formed using the effective channel matrices⁴\(\mathbf {G}_{c(i),\pi _{c(i),k}^{j}} = \tilde {\mathbf {H}}_{c(i),\pi _{c(i),k}^{j}}\mathbf {V}_{k-1}^{0}\). The user selection within a cluster is performed by using a SAS algorithm similar to what we proposed in [35] and described in Algorithm 3 therein. For ease of reference, the pseudocode of the SAS algorithm is described in Algorithm 1 here. The main difference between [35] and here is that the solution values s _x and \(s_{\hat {\mathbf {x}}}\) are now achievable weighted sum rates as per (6). The ordered user vectors x and \(\hat {\mathbf {x}}\) are used for \(\pi _{c(i)}^{j}\) in (6). The rest of the operation of the algorithm is unchanged. The SAS algorithm operates in parallel separately for each cluster.

Two positive variables B₁ and B₂ limit the iterations and control how closely the algorithm approaches the optimal solution with the trade-off of the algorithm’s performance and its acceptable complexity. The larger the values of B₁ and B₂, the closer the algorithm comes to the optimal solution, but the computational complexity also increases as the algorithm iterates longer. The SAS algorithm starts with the variable τ _t (analogous to the temperature in metallurgical annealing) equal to τ_hot. It continues until τ _t “cools” to the value of \(\phantom {\dot {i}\!}\tau _{f} = \tau _{\text {hot}} \phi ^{B_{1}}\). ϕ is a uniformly distributed random variable in the interval (0,1). The neighborhood function N(x(n)) at random either deletes a random user from the vector x(n), adds a random unscheduled user (if possible), replaces one scheduled user with a random unscheduled one, or swaps the encoding order position of two random scheduled users. Each of these actions has an equal probability of being chosen.

We refer the reader to [35] for more details on our SAS algorithm. Note that no change to the scheduling algorithm is required for rotating clustering; better cluster patterns for users are automatically detected by the algorithm through the corresponding more favorable channel gains and/or achievable rates during that pattern, making the users more likely to be scheduled during those better patterns.

3.1 Simulation results for six-cell layout

The clustering pattern is changed every T _cl scheduling intervals, and we set t _c =100 for the size of the averaging window in the PF metric. We consider a minimum of four full sets of rotations through both patterns for our simulations (with that minimum occurring at the slowest rotation speed). The cumulative distribution functions (CDFs) of the average achievable rates per user for K _c =12 and various T _cl values are depicted in Fig. 3. As this figure shows, rotating clustering has the most benefit in particular for those users with poorer channel gains (e.g., those with an average achievable rate around the 5th percentile). In comparison, users with better channels (who achieve higher average rates, such as those around the 90th percentile) see their average achievable rate drop with rotating clustering. The poorer 5th percentile users are those having low SINRs in a cluster in one of the patterns. Their achievable rates significantly benefit from rotation because they are located in a position in the cluster where they can have a better SINR after the rotation. The better users trade off their average achievable rate to provide higher overall fairness to the system.

Considering different rotation speeds, Fig. 3 demonstrates that the rotation rate has an impact on each user’s average rate. This is the most easily seen for lower-rate users like those around the 5th percentile. The 5th percentile per-user rates generally increase with decreasing T _cl , but eventually reach an upper limit, where even faster rotation yields no further significant gains. With PF scheduling, the gain in 5th percentile rate over that achieved with fixed clustering ranges between 23.5% to 28%. The gains in rates per user with PF scheduling are further emphasized in Fig. 4, which depicts, for each percentile of the average per-user rate, the percentage gain in average achievable per-user rates obtained when using rotating clustering, relative to the rates achieved with fixed clustering. As can be seen, the majority of the achieved per-user rates (those up to about the 65th to 72nd percentile) experience a gain with rotating clustering, at the expense of the rest. However, even at those remaining upper percentiles, the relative loss in rate is much smaller than the relative gain for the lower percentiles; at worst, the upper percentile rates drop by about 5%.

In each cluster, there are some users with particularly poor channel gains and even cluster rotation cannot help improve them much (such as those users located near the cell borders). These users have a very low chance to be selected by the MT scheduler; their average achieved rate is very close to (and sometimes equal to) zero, especially for fixed clustering. Cluster rotation helps these users by occasionally giving them better channels; however, there still remain some users that suffer from starvation. Nonetheless, although the probability of starvation is still non-zero for MT scheduling, it is reduced significantly (by a factor of over half for K _c =12) by using cluster rotation, as seen in Fig. 3.

With rotation, the best users’ average achievable rates are reduced because they may be scheduled less often; however, they achieve higher instantaneous rates when they are selected. As an example, consider a theoretical extreme case of maximum throughput scheduling, where with fixed clustering, a few users would be scheduled most of the time with high throughput. With rotating clustering, instead about twice as many users would be scheduled overall, half of this total in each clustering pattern. As they would be scheduled about half as often (mostly only during their favorable pattern), their average achievable rate would also be about halved. However, they would achieve somewhat better instantaneous rates when scheduled, leading to an increased (albeit slightly) achievable sum rate. Similar and larger effects on the instantaneous and average achievable rates are seen with PF scheduling. As expected, while the average sum rate of PF is less than that of maximum throughput scheduling, its 5th percentile average achievable per-user rates are higher than for maximum throughput, and there is less overall variation in the average per-user rates achieved.

Figure 5 shows the average achievable sum rate for maximum throughput and proportionally fair scheduling vs. K _c over the area of a macrocell in an arbitrary cluster. In Fig. 5, T _cl =100, which is equal to t _c . As seen, rotating clustering outperforms fixed clustering for both scheduling metrics. However, rotating clustering increases the average achievable sum rate with PF relatively more than with maximum throughput. For instance, for MT scheduling, rotating clustering provides slightly higher throughput, increasing about 0.85% and 0.75% for K _c =4 and K _c =12, respectively, while for PF scheduling, the throughput is more significantly higher (about 1.4% and 3.2%, respectively, for K _c =4 and 12). This is expected because the users that take the most advantage of the rotation are the users either near the border of the cluster or in poor coverage areas, who are scheduled more often with PF than with maximum throughput.

In Fig. 6, the average achievable sum rate vs. K _c for proportionally fair scheduling considering different rotation speeds is presented. As seen, while rotating clustering still outperforms fixed clustering in terms of sum rate, faster rotation (i.e., smaller T _cl ) yields diminishing gains. For instance, rotating clustering provides higher throughput with respect to fixed clustering with faster rotation⁸, increasing by about 2.1%, 3.2%, 4.1%, and 4.5% with K _c =12 for T _cl =200, 100, 50, and 25, respectively. Considering an upper limit seen in the sum rate, rotating faster would not provide any further significant increase, but would increase complexity in signaling overhead for cluster setup. For our system, the limit is reached at about T _cl =50 (half of t _c ). Smaller values of T _cl (e.g., T _cl =25) yield almost no additional sum rate; the additional gain in sum rate relative to fixed clustering is only a few tenths of a percentage.

Assuming the scheduler has selected a user with poor channel gains in a cluster, the chance of this user being chosen again by the scheduler is comparatively low since its priority weight will have likely dropped in accordance with the update of its average rate by the exponential filter. The user’s priority will gradually increase as time passes if the user is not scheduled. If the cluster pattern change interval is smaller than t _c , the possibility of that user being in a better position in another cluster pattern, and consequently dramatically improving its priority by virtue of its higher SINR and thus achievable rate, is increased. Similarly, a user with high SINR in a cluster that finds itself near a cluster edge after the pattern rotates will have its priority suddenly drop. It will be less likely to be scheduled until either the PF scheduling window passes or the cluster pattern rotates back, whichever comes first. Hence, faster rotations relative to t _c can potentially result in higher priority weights and higher sum rates. However, decreasing T _cl does not imply a linear increase in priority weights.

The upper limit on the performance with increasing rotation speed is understandable. In the PF scheduler, there are two factors influencing whether a user is scheduled: its potential instantaneous rate and its weight. The former is largely influenced by the cluster rotation, whereas the latter, among other things, is an indicator of how long the user has gone without being scheduled. The weight is updated with an exponential filter, meaning its value increases exponentially with time. Specifically, the increase is determined relative to the time constant (averaging window) t _c , expressed as a number of scheduling intervals. If a user is not scheduled for ν scheduling intervals, its weight will increase by a factor of roughly exp(ν/t _c ). If the clustering pattern rotation interval is small relative to t _c , the conditions for a user will become favorable or unfavorable sooner than the time it takes for the weight to change significantly. In other words, when T _cl is a fraction of t _c , the weights change little between the instances in which a given user is scheduled. Hence, the specific value of T _cl is no longer much of a factor, leading to the upper limit on (saturation of) the performance with increasing rotation speed.

With larger K _c , cluster rotation is more effective at increasing the sum rate; for example, the achievable gains with T _cl =100 are 1.4% and 3.2% for K _c =4 and 12, respectively, using PF scheduling. Faster rotation with larger K _c is more beneficial. For instance, with K _c =8 and comparing the rotation speed of T _cl =200 to T _cl =100 and 50, the sum rate gains are 0.9% and 1.4%, respectively, whereas with K _c =16, the gain increments are 1.4% and 2.4%. The probability of having users that are located in the cluster with favorable channel gains is increased with larger K _c , simply as a result of multiuser diversity. Thus, being able to schedule more of these users improves the sum rate. Recall, though, that users are uniformly distributed over the entire coverage area. This statistically results in a higher proportion of users who are farther from a BS (with poorer channel gains) than those who are nearer. An increase in K _c thus also means an increase in the total number of “poorer” users. As was seen earlier, it is those users who benefit the most from rotation, explaining further why the gains with rotation are better with higher K _c .

An increase in K _c also means more delays in scheduling users, as a larger pool competes for the same limited resources. Faster rotation also means less of a wait for any given user’s cluster pattern to be at its best, and thus potentially for the user to be scheduled, improving its average rate. In other words, faster rotation somewhat compensates for the increased scheduling delays as K _c grows, thus leading to the higher gains in rate observed with faster rotation at larger K _c .

Figures 7 and 8 further demonstrate the effects of rotating clustering (with T _cl =100) by depicting the average achievable rate per user based on the user’s position within the cellular network; Fig. 7 depicts MT scheduling, whereas Fig. 8 depicts PF. Figures 7a and 8a show the average achievable rates with fixed clustering. The highest rates are unsurprisingly achieved by users nearest to a BS, especially macro BSs, but also to a lesser extent pico BSs. The lowest achievable rates (darkest red) are seen near any macrocell corner where three clusters meet. We will henceforth refer to these areas as the “corner” areas for shorthand. Lower rates are also seen in the area around the border of each cell, whereas somewhat higher achievable rates (orange to yellow to white) are seen near areas that correspond to the directions of the macro BS boresights. For any macro BS, users located near the left or right borders of the cell (i.e., at a 30° angle relative to the antenna boresight) receive weaker signals on account of the antenna directivity pattern. This situation is at its worst near the corners of the macrocells, at the furthest distance from the site. Furthermore, although the corners are surrounded by picos, the received signals from those picos are also weak due to being beyond the picocell borders. Hence, even with clustering (either fixed or rotating) and coordination, users in the corners achieve consistently smaller user rates, compared to the users in other locations of the cluster.

Figures 7b and 8b show the achievable rates with rotating clustering. In comparison, the average rates (or the colors) are more evenly distributed over the entire area, indicating higher overall fairness. As expected, the largest increase in average achievable rate is experienced by users that are closest to the cluster borders in the fixed scheme. Those users previously received the worst SINRs and/or the most ICI. In contrast, most users who are located close to the macro BSs or near the directions of the macro boresights receive very good signal power. The largest decrease in average achievable rates relative to the fixed scheme is by that latter group of users, as well as to a lesser extent users nearby pico BSs. It is these latter users who trade off some of their average rates to provide more fairness and uniformity of throughput across the coverage area. It is interesting to note that there are fairly large areas where the average achievable rate of a user does not change much with rotating clustering. These areas and users are those located in the interior of the cluster. Their conditions (in terms of both useful signals and interference) are more or less the same under either cluster pattern. Hence, rotation does not change their situation much, and so, their achievable rates do not change appreciably from the fixed scheme either. There is also little difference in the rates seen in the cell corner areas with rotation. The rather small changes that exist there are not readily visible in Figs. 7 and 8. However, where three clusters meet in the fixed scheme, those corners do experience a small boost in rates with rotation due to periodically increased coordination. Likewise, the corners at the cluster centers in the fixed scheme see a small decrease in rates with rotation, as signals received there are no longer always coordinated from all nearby BSs.

3.2 Simulation results for three-cell layout

Similarly to the previous section, the clustering pattern changes (rotates) every T _cl scheduling intervals. For the six-cell layout, as described in Section 3.1, we simulate for a minimum of four complete sets of rotations. Assuming T _cl =200, which corresponds to the slowest (non-zero) rotation speed, the total number of simulated channel realizations (or scheduling intervals) for every drop of users is thus 1600. To compare the three-cell layout with the six-cell layout, we consider two different scenarios. First, we keep the same total number of 1600 channel realizations per drop with four complete sets of rotations between the five clustering patterns. We also maintain the longest pattern duration as 2t _c . Thus, t _c for the PF window becomes 1600 realizations/drop ÷4 sets/drop ÷5 patterns/set ÷2 windows/pattern =40 (realizations per window). This gives us different rotation speeds corresponding to pattern durations of T _cl =[10,20,40,80]. Secondly, we keep the value of t _c equal to what we use in Section 3.1, i.e., 100, and the same pattern durations, i.e., T _cl =[25,50,100,200]. With still four full sets of rotations at minimum, this results in 4000 channel realizations per drop. The simulations with these two scenarios were run for K _c =12, and the results were compared with each other. We observed that the results of both scenarios were very close to each other⁹. Thus, we present only the results for where the number of channel realizations per drop equals¹⁰ 1600 and t _c =40.

In Fig. 9, the CDF of the average achievable rates per user for K _c = 12 is illustrated. Similarly to the results depicted by Fig. 3, rotating clustering again yields higher per-user rates for those who are located near cluster edges or in otherwise poor coverage areas in the three-cell layout (e.g., the users with rates around the 5th percentile). Comparison of these two figures is also interesting. As seen in Fig. 3 with MT scheduling, while rotating clustering improves the per-user rates in the six-cell layout, there are still some users whose achievable rates are very low, even to the point of starvation. However, cluster rotation in the three-cell layout improves the average throughput of users considerably such that the probability of very low rates and/or starvation becomes almost zero. There is a larger trade-off in the high-percentile users’ rates to achieve this, though.

In the three-cell layout and with PF scheduling, the rotation speed corresponding to T _cl =t _c yields a per-user rate increase of about 0.5 bits/s/Hz at the 5th percentile, while the improvement with the six-cell layout is around two thirds of that. In relative terms, the 5th-percentile rate improvements with rotating clustering (at T _cl =t _c ) over fixed in the six- and three-cell layouts are, respectively, about 25.7% and 47.3%. This indicates cluster rotation is even more effective at improving the cluster-edge user rates in the three-cell layout. Considering the faster rotation speed corresponding to T _cl =t _c /2, the additional 5th-percentile users’ rate increase (relative to T _cl =t _c ) is about 1.3% in the six-cell layout, while this increase is about 3.7% using the three-cell layout. This shows that faster rotation in the three-cell layout is more beneficial than in the six-cell layout to help cluster-edge users to gain further higher throughput. This result is not surprising, considering the former rotates through a larger set of patterns.

We also investigate the user rates vs. their position within the network in Fig. 10 (MT scheduling) and Fig. 11 (PF scheduling). Basically, the influence of cluster rotation on achievable per-user rates with MT and PF scheduling is similar to what we have discussed in Section 3.1. As before, it is users located at the cell corners away from the sites that achieve the lowest rates. To a lesser extent, users near the middle of cells, not especially near a macro BS nor a pico BS, also achieve somewhat lower rates. As is obvious from Figs. 10 and 11, the achieved users’ rates are distributed more uniformly with rotating clustering than with fixed clustering for both MT and PF scheduling. Furthermore, the overall rates are increased. The achievable throughput of those users who are located in the corners of a cell is improved, and the size of the areas experiencing lower rates is diminished. This is particularly noticeable in Fig. 10 for MT scheduling; the colors in the cell corners increase from dark red to orange.

Cell-center users are also significantly affected by cluster rotation. Note that in fixed clustering, and in the patterns in Fig. 2a, c, and e, two parts of the cluster come from the same site. Because of the macro BS antenna pattern, the coverage of the beams from those two BSs has little overlap. Hence, those two cells are essentially impacted by, at best, two coordinated signals: one from their own cell and one from the third BS at the other site that contributes to the cluster. However, with rotation through the patterns in Fig. 2b, d, receiving a significant coordinated signal from three BSs is more common. Thus, users in those two cells in the cluster see the most improvement from rotation¹¹. The trade-off now comes from users in the third cell of the fixed scheme, who with cluster rotation now are forced to experience less advantageous cluster patterns for the sake of overall network fairness.

Figure 12 demonstrates the achievable sum rates vs. K _c for the three-cell layout. Much like the six-cell layout, MT scheduling again does not display a gain in sum rate from cluster rotation, for similar reasons as described in Section 3.1. However, the throughput with PF scheduling does again increase considerably. For example, with K _c = 12, comparing the achievable throughput of rotating vs. fixed clustering shows an increase of 4.5% and 6.7%, for T _cl equal to t _c and t _c /2, respectively. Comparing the six- and three-cell layouts with the same number of users in each cluster, the six-cell layout generally yields higher values for area spectral efficiency on average. For example, the average total sum rates per macrocell for K _c =12 are 8.8 and 13.6 bits/s/Hz, respectively, for the six- and three-cell layouts with PF using fixed clustering. Yet, also note that the macrocell area in the six-cell layout is half of that in the three-cell layout (see Endnote 5). Hence, the sum rate per unit area (e.g., per km²) is larger in the six-cell layout. However, the per-user rate improvement achieved with cluster rotation in the three-cell layout is much more significant. Faster rotation yields higher gains in this layout than in the six-cell layout. For instance, the additional sum rate improvement achieved by setting T _cl =t _c /2 relative to T _cl =t _c is 2.1% in the three-cell layout, while this gain in the six-cell layout is only 0.8%. Although faster rotation improves the average sum rate, the upper limit for higher speeds of rotation is still similar to what we described in Section 3.1. Thus, we only present the results for T _cl equal to t _c and t _c /2. All these comparisons demonstrate that the rotating cluster method overall helps the system performance in the clover-leaf-shaped network layout more than in the hexagonal-shaped layout.

3.3 Comparison of rotating cluster method with dynamic cluster method

In [30], a dynamic clustering method was proposed and compared with static clustering. There are several differences in the simulation methodology in that paper compared to our own (e.g., a single-tier homogeneous network, differing numbers of users and antennas, the precoding method). However, it can still serve as a rough guide for the performance benefits of a fully dynamic scheme over static clustering and that can be used for comparison with our less complex rotating clustering method. The simulations in [30] assumed three macrocells per site, four transmit antennas per macrocell, two antennas per user, and PF scheduling. The dynamic scheme was shown to have approximately an 18% gain in average sum rate per cell (Fig. 5 in [30]) vs. a static clustering scheme of three cells per cluster with one macrocell contributing from each site (similar to the macrocells in our three-cell layout). The gain in 5th percentile user rate was about 22% (Fig. 6 in [30]); this gain approximately doubled when the maximum cluster size of the dynamic scheme increased from three to six, though at the same time also enforcing that just a single data stream be sent to each user (Fig. 9 in [30]).

In comparison, for our six-cell layout, the 5th percentile user rates for K _c = 12 and PF scheduling increase by about 23.5% to 28% over static clustering for the minimum and maximum examined rotation speeds. In the three-cell layout, the gains were even higher, i.e., 47–53% for K _c = 12. The gain in sum rate per cell was lower though, about 2.1–4.5% for K _c = 12 in the six-cell layout, and about 4.5–6.7% for K _c = 12 in the three-cell layout. However, we remind the reader that our rotating cluster scheme was primarily designed to help cluster-edge users (e.g., those around the 5th percentile); any gains in cell sum rate are an added secondary bonus. Hence, we can see that, compared to the fully dynamic scheme of [30], our rotating scheme provides gains for cluster-edge users that are on par with or better than those in [30], but at the trade-off of lower gains in the sum rate per cell.