You are on page 1of 15

Wireless Data Multicasting with Switched Beamforming Antennas

Honghai Zhang*, Yuanxi Jiang* , Karthik Sundaresan*, Sampath Rangarajan*, Baohua Zhao *Mobile Communications and Networking Research, NEC Laboratories America Dept. of Computer Science and Technology, University of Science and Technology of China

AbstractUsing beamforming antennas to improve wireless multicast transmissions has received considerable attention recently. The work in [20] proposes to partition all single-lobe beams into groups and to form composite multi-lobe beam patterns to transmit multicast trafc. Depending on how the power is split among the individual beams constituting a composite beam pattern, two power models are considered: (i) equal power split (EQP), and (ii) asymmetric power split (ASP). This work revisits the key challenge - beam partitioning in the beamforming-multicast problem considered in [20] and makes signicant progress in both algorithmic and analytic aspects of the problem. Under EQP, we propose a low-complexity optimal algorithm based on dynamic programming. Under ASP, we prove that it is NP-hard to have ( 3 )-approximation algorithm for any > 0. 2 For discrete rate functions under ASP, we develop an 3 APTAS, an asymptotic ( 2 + )-approximation solution (where 0 depends on the wireless technology), and an asymptotic 2-approximation solution to the problem by relating the problem to a generalized version of the bin-packing problem. For continuous rate functions under ASP, we develop sufcient conditions under which the optimal number of composite beams is 1, K, and arbitrary, respectively, where K is the total number of singlelobe beams. Both experimental results and simulations based on real-world channel measurements corroborate our analytical results by showing signicant improvement compared to state of the art algorithms.

I. I NTRODUCTION Designing efcient link layer multicast solutions is becoming increasingly important in data dissemination for group communications (such as mobile TV, Sports Telecast, Video Teleconference, etc.). While the shared, broadcast nature of the wireless medium provides natural support for wireless multicast services, the multicast transmission rate is limited by the client with the worst channel conditions in the group (e.g., [3]). Beamforming antennas, by virtue of their ability to focus energy in a specic direction, provide a natural solution to improve

the received signal strength at the weakest client, which can potentially improve multicast performance. However, it is challenging to apply beamforming technologies to multicast transmissions because of the inherent tradeoff between multicasting and beamforming. While beamforming increases the signal energy in a particular direction, it also reduces the energy in other directions, thereby restricting the wireless broadcast advantage, which is a key component in multicasting. Recently, Sen et al. [17] considered the problem of integrating multicast with beamforming and proposed to transmit with an omni-directional beam rst, followed by one or several sequential single-lobe transmissions. Several other works [1], [12], [20] pointed out that in a strong line-of-sight (LOS) environment, such as indoor channels at 60 GHz or outdoor wireless systems, the beamforming gain is signicant, especially when the number of antenna elements is large (e.g., Fidelity Comtech [6] provides antennas with eight elements and a typical 60 GHz system allows 32-64 antenna elements [9], [23]). With a large number of antenna elements in a LOS environment, the antenna can form very narrow single-lobe beams that can be roughly viewed as nonoverlapping [20]. In other words, the received energy at any location from one particular single-lobe beam dominates that from all other single-lobe beams. Under such a context, Sundaresan et al. [20] showed that composite multi-lobe beam patterns are needed to address the multicast-beamforming tradeoff efciently. [20] formulated the problem as minimizing the aggregate transmission time in disseminating a common message. This is achieved by partitioning single-lobe beams into multiple groups and forming a composite beam for each group with sequential transmission on each of the composite beams. Depending on how the power is split among the individual beams constituting a composite beam pattern, two models are considered: (i) equal power split (EQP), and (ii) asymmetric power split (ASP). The

key challenge is how to partition the beams into groups. Several algorithms are reported in [20] to address the challenge. In this work, we revisit the key challenge of beam partitioning considered in [20] and make signicant progress on this problem. We accomplish this through rigorous complexity analysis and designing new lowcomplexity algorithms with performance guarantees for both EQP and ASP models. Our main contributions can be summarized as follows.
Under the EQP model, we provide a lowcomplexity, dynamic-programming-based optimal solution for both continuous and discrete rate functions. The complexity of our algorithm is O(K 2 ), in contrast to the O(K 7 ) complexity of the optimal solution in [20], where K is the total number of non-overlapping single-lobe beams. Under the harder ASP model, currently there exists no hardness results or approximation solutions. We present several key results in this context. 1. We prove that it is NP-hard to have ( 3 )2 approximation solution for a general rate function for any > 0. 2. For discrete rate functions, we show that multicastbeamforming problem can be converted to a generalized version of the bin-packing problem. This allows us to leverage and apply generalized-binpacking algorithms to obtain an APTAS (Asymptotically Polynomial Time Approximation Scheme) 1 as well as an asymptotic ( 3 + )-approximation 2 solution for the multicast-beamforming problem, where 0 depends on the discrete rate function used by the wireless technology. We also develop a novel asymptotic 2-approximation solution that applies to all discrete rate functions. In retrospect, this also yields an asymptotic 2-approximation solution for the generalized bin-packing problem, which is of independent interest. 3. For continuous rate functions, we derive generic sufcient conditions for the rate functions under which it is optimal to have (i) one, (ii) K , and (iii) arbitrary number of composite beams, respectively. In particular, we show that if the rate is a nondecreasing concave function of SNR, it is optimal to have one composite beam containing all singlelobe beams, which coincides with the result in [20] for the special case of Shannon-capacity rate
For the denition of APTAS and approximation algorithms, please refer to Appendix A and [15].
1

function. To corroborate the theoretical analysis, we evaluate the algorithms based on real traces from signal measurements of an eight-element phased array antenna in an outdoor testbed, as well as real-world outdoor experiments. Comprehensive evaluations indicate that the proposed algorithms signicantly improve the state of the art in literature. The multicast delay reduction for 802.11a and 802.11b is up to 20% and 25%, respectively, compared to the algorithms in [20] under the ASP model. The rest of the paper is organized as follows. In Section II, we discuss the background and the optimization framework. The proposed algorithms are presented in Sections III and IV for EQP and ASP models, respectively. Evaluation of the proposed solutions based on real-world traces is presented in Section V and the experimental results are presented in VI. We discuss the related work in Section VII, followed by conclusion in Section VIII. II. BACKGROUND
AND OPTIMIZATION FRAMEWORK

A. Background and motivation A smart antenna system combines multiple antenna elements in an array with signal processing capability to optimize its transmission and/or reception pattern. In a beamforming antenna system, each antenna element can be pre-coded with a complex weight, forming a beam pattern, such that the total energy of all beams along a certain direction in the physical or signal space is maximized. Beamforming can be either adaptive, or switched. The former generates the pre-coding weights dynamically based on the receiver channel conditions and the latter provides a set of pre-computed beams to be used at any time instant. Although adaptive beamforming provides better antenna gain, it also requires sophisticated signal processing capability and complexvalued channel feedback. On the other hand, switched beamforming achieves better performance-complexity tradeoff and is thus considered in this work. In a switched beamforming system, the antenna typically provides a set of K single-lobe beam patterns of degree 360/K covering the entire azimuth of 360o where K is the number of antenna elements in the array. In environments with strong LOS (line-of-sight) such as in-door channels at 60 GHz or out-door wireless systems, at any given client location, the power received from one single-lobe beam (that is closest to the line-of-sight to the client) typically dominates that from all other single-lobe beams. This work considers such line-of-sight scenarios and assumes that the received energy from all other 2

single-lobe beams are negligible compared to that from the beam with the strongest signal. Under this assumption, Sundaresan et al. [20] proposed to partition the single-lobe beams into groups, with each group of beams forming a composite beam, and to transmit on each composite beam. The key challenge is then to determine the optimal partitioning of beams into groups. Several algorithms were presented in [20] to solve the problem. This work makes signicant progress over [20] on both the algorithmic and analytic aspects of the beam partitioning problem. While this work focuses on the non-overlapping beam patterns, we consider the overlapping beam patterns for video delivery in a parallel work [25]. B. Network model We consider a single-cell environment where an AP with a smart antenna system serves multiple clients using link layer multicast. All clients measure the SNR values from each beam and feedback the best SNR and the index of the best beam back to the AP. Assume that yi is the SNR value of client i when it is served by the AP with a single best beam pattern and bi is the best beam of client i. Let Ck denote all clients who are best served by beam k, i.e., Ck = {i : bi = k}. Dene the effective SNR of a beam k as k = min{yi : i Ck } representing the minimum SNR among all clients served by beam k. When the AP transmits at a rate that corresponds to k using a single beam pattern k, all clients associated with beam k can decode the packet. Note that the usage of all K beams (at appropriate rate) will cover all clients. C. Optimization framework We denote R() as a general non-decreasing rate function of an SNR value . Assume that there are totally K switched beams and N users in the system, and the multicast data size is L bytes. The objective is to partition the beams into G groups and transmit L bytes sequentially on each beam group, such that the aggregate transmission delay to deliver L bytes to all clients is minimized. Assuming that there is a switching delay W for each transmission to a beam group, the objective can be written as
G

beams. Depending on how the power is split among multiple beams in a given group, two possible models are considered (as in [20]): EQP (EQual Power) model and ASP (ASymmetric Power) model. III. EQP M ODEL A. Problem formulation Under the EQP model, power is equally split among multiple beams. While this method of power allocation is not optimal, it is a simple, yet reasonable choice. Let Bg denote the set of beams of group g. Due to equal power splitting, the SNR value of each beam in group g reduces by a factor of |Bg |. In order that all clients served by a group of beams can decode the packet, the AP transmits at a rate corresponding to the lowest SNR value among all beams in the group. Therefore, the effective e SNR value for the group g under EQP model is g = minkBg {k }/|Bg |. Now the objective in Eq. (12) can be written as
G

min
g=1

(W +

L ). R(minkBg {k }/|Bg |)

(2)

The number of partitions G and the partitions Bg (g = 1, , G) are the variables to be optimized. Sundaresan et al. [20] showed that this problem under the specic Shannon capacity rate function (i.e., R() = log(1 + )) can be solved, albeit at a high complexity of O(K 7 ). In the following, we develop an optimal solution to problem (2) via dynamic programming with a low complexity of O(K 2 ) and it is applicable to any nondecreasing rate function R(). B. Optimality Denote T (P ) as the total transmission time of partition P . If P contains only one group Bg , its transmission time (including delay between switching beams) is
T (Bg ) = W + L . R(minkBg {k }/|Bg |)

(3)

min
g=1

(W +

L ) e R(g )

(1)

e where g is the effective SNR value of group g (to be decided). When the AP transmits on a group of beams simultaneously, the transmit power on each beam decreases because the net power is split among multiple

Assume that the rate function R() is non-decreasing with respect to the SNR value . Let be a list of all beams sorted in the decreasing (or increasing) order of their effective SNR (i.e., k ). Now, we can establish the following lemma. nc Lemma 1: If P is a non-contiguous partition of nc resulting in a total transmission time T (P ), there c of such that T (P c ) exists a contiguous partition P nc T (P ). 3

This lemma generalizes the Lemma 1 in [20] where the Shannon capacity rate function is assumed. The proof is similar to that in [20] and is omitted. This lemma reduces problem (2) to a contiguous partition problem, where we rst sort the beams in the decreasing or increasing order of their effective SNR (i.e., k ), and then divide the ordered set into groups consisting of contiguous elements so as to minimize the total transmission time. We will now show that the optimal solution to our problem can be constructed recursively from the optimal solutions to its sub-problems, thereby enabling a dynamic-programming based solution. Theorem 1: Let be a list of all the beams sorted in the decreasing (or increasing) order of their effective SNR. If P is an optimal contiguous partition of (i.e., with the minimum transmission time) and P = (P\BG , BG ), where BG is the last (most recent) group of beams determined, then P\BG is an optimal partition for the set of beams in \BG (which denotes the set of beams in but not in BG ). Proof: We will prove this by contradiction. If P\BG is not an optimal partition for the set of beams in \BG , we can construct an optimal partition P\BG for it. Denote the new partition of as P = (P\BG , BG ). is Now the total transmission time of P
T (P ) = T (P\BG ) + T (BG ) < T (P\BG ) + T (BG ) = T (P ).

The initial condition is


S1 = T ({1 })

(5)

The complete algorithm (DP-EQP) based on dynamicprogramming is illustrated in Algorithm 1. Algorithm 1 DP-EQP:Dynamic-programming-based algorithm for EQP 1: Sort the beams in the decreasing order of their SNR. Denote the resulting permutation as (1 , 2 , , K ). 2: Compute S1 using Eq. (5) 3: for k = 2 to K do 4: Compute Sk using the recursive equation (4). 5: end for 6: Return the optimal multicast transmission time SK and the optimal partition. Complexity of DP-EQP: Step 1 of the algorithm involves sorting and can be computed with O(K log K) complexity, while steps 3-5 require O(K 2 ) complexity. Therefore, the total complexity of the algorithm DP-EQP is O(K 2 ). Remarks: 1) In order to return the optimal beam partitioning, DP-EQP needs to keep track of the intermediate optimal state at each step 4, or use back-tracking after obtaining the optimal cost. Both approaches are standard methods in dynamic-programming and are omitted. 2) DP-EQP can be applied to arbitrary rate functions R, including both continuous and discrete functions, as long as they are non-decreasing with respect to the SNR values. IV. ASP M ODEL A. Problem formulation Under the ASP model, power can be optimally allocated among multiple beams in a group such that the minimum SNR value across all beams in the group is maximized. In [20], the authors showed that the optimal power allocation to beam k within a group is given by k = k , where
= 1
1 kBg k

This leads to a contradiction, given that P is an optimal partition. C. Optimal Dynamic Programming Solution: DP-EQP Based on the optimality principle in Theorem 1, we design the following dynamic-programming-based approach to compute the optimal solution. Assume that = (1 , 2 , , K ) is the complete list of beams sorted in the decreasing order of their effective SNRs. Denote Sk as the total transmission time of the optimal partition of the rst k beams in . From Theorem 1, Sk can be recursively computed as
Sk = min (Sj1 + T ({j , , k }))
1jk

(4)

(6)

where T ({j+1 , , k }) is the transmission delay of the last group and is calculated using Eq. (3). Since the beams are sorted in the decreasing order of their SNRs, the computation can be simplied as
L T ({j+1 , , k }) = W + . R(k /(k j))

e is the effective SNR (i.e., g ) of group g. Under this model, the objective in (12) can be written as G

min
g=1

(W +

L R(1/

1 ). kBg k )

(7)

Again, the optimization variables are the number of groups G and the set of beams (i.e., Bg ) in each group g. B. Hardness of the Problem We rst prove that problem (7) is NP-hard for a general rate function. Theorem 2: There is no approximation algorithm with a guarantee of 3/2 - for Problem (7) for > 0 unless P = NP. Proof: It was shown in [21] that it is NP-hard to 3 have a ( 2 )-approximation algorithm for the binpacking problem. We reduce the bin-packing problem to a special case of problem (7). Bin packing problem: Given n items with sizes s1 , s2 , sn (0, 1], nd a packing in unit-sized bins that minimizes the number of bins used. Consider the following special case of problem (7). There are n beams with the effective SNR of beam i being i = 1/si . Further, let L = 1 and the switching delay W = 0. Let the rate function be
R() = 1, 0, 1 otherwise

C. Discrete Rate Function In this subsection, we consider problem (7) with discrete rate functions. As the general beam multicasting problem is NP-hard, we turn to approximation solutions to solve the problem. A general discrete rate function R() can be represented using a step function as,
R() = Rm RM if m < m+1 , for 1 m < M if M . (9)

where m typically represents a Modulation Coding Scheme (MCS), and m , Rm are the SNR threshold and the transmission rate with MCS mode m. We assume that the effective SNR (k ) of each beam k is at least 1 (otherwise, some users under beam k cannot be served with any MCS by any beam, and have to be dropped from consideration). For such discrete rate functions, we can convert the beam partition problem into the following generalizedcost variable-sized bin packing (GCVS-BP) problem [7]. Mapping to Bin Packing: GCVS-BP problem: We are given L types of bins, each with innite supply. A bin of type l has size 0 < bl 1 and cost cl > 0. Without loss of generality, we assume that b1 b2 bL . Items of sizes in (0, b1 ] are to be partitioned into J subsets. Each subset j of items are then packed into a bin of type lj such that the size of the bin type lj is at least as large as the total size of all items in the subset j . The total cost of the packing is J clj . The goal is to nd j=1 a feasible packing such that the total cost is minimized. We next convert beam partitioning problem (7) under ASP to the GCVS-BP problem. For each beam pattern i with effective SNR i , we construct an item with size 1/i . For each MCS mode m, we create a bin type L with size bm = 1/m and cost cm = W + Rm . There are totally M types of bins. Each group Bg of beams corresponds to a bin. If the effective SNR (dened in Eq. (6)) of the group of beams satises m(g) , it can choose transmission rate Rm(g) , which corresponds to packing the items derived from Bg to a bin of type L m with a cost of W + Rm(g) (as m(g) implies 1 iBg i 1/m(g) = bm(g) , corresponding to the binsize constraint). The resulting total cost of all bins is
G

(8)

First note that the optimal solution to this problem takes a nite value because if we let each beam occupy one group, the resulting partition has a nite cost. Therefore, for each group g in the optimal partition, its effective SNR is at least 1 (so that the resulting cost is nite). We next establish a one-to-one relationship between a solution for the bin-packing problem and that for our beam partitioning problem. For each non-empty bin Bj in the bin-packing solution, we can construct a group of beams in our problem, where each beam i in the group corresponds to the item i in the bin. Since the total size of all items in a bin cannot exceed 1, i.e., iBj si 1, this indicates that the effective SNR of the corresponding beam group (which is equal to 1/( iBj 1/i ) = 1/( iBj si )) is greater than one. As a result, the cost of the corresponding beam group is 1 from Eq. (8). Therefore, the objective value of any nite solution to our problem is equal to the number of non-empty bins in the bin-packing solution. Clearly, the mapping between the bin-packing solution and the solution to the beam partition problem is one-to-one and the transformation can be done in polynomial time, which completes the proof. We next consider two cases depending on whether the rate function R() is discrete or continuous. 5

W+
g=1

L Rm(g)

which is exactly the aggregate transmission delay. Table I shows the mapping from the beam multicast problem to the GCVS-BP problem. With the above mapping, we can apply algorithms developed for the GCVS-BP problem to solve the beam

TABLE I M APPING BETWEEN A BEAM MULTICAST PROBLEM TO A GCVS-BP PROBLEM Variables in beam multicast beam SNR i SNR threshold m of MCS m Rate Rm of MCS m P Total delay G (W + R L ) g=1
m(g)

and not amenable to implementation. Hence, more efcient solutions are desired.
3 An Asymptotic ( 2 + )-approximation Algorithm: Kang and Park [11] studied the GCVS-BP problem with variable cost functions satisfying ci cj for any bi < bj , (10) bi bj

Variables in GCVS-BP object size si = 1/i bin size bm = 1/m bin cost cm = W + L/Rm P Total cost G (W + R L ) g=1
m(g)

Algorithm 2 General algorithm for beam multicasting 1: Convert the beam multicast problem (7) to a GCVSBP problem with item sizes 1/i , i = 1, , N , bin sizes 1/m , m = 1, , M and bin costs W + L , m = 1, , M . Rm 2: Apply a generalized bin-packing algorithm to solve the converted GCVS-BP problem. 3: Map the items in each bin in the bin-packing solution to a group of beams for simultaneous transmission to obtain a beam multicast solution.

and proposed an algorithm with asymptotic performance ratio of 3 . They show that their proposed algorithm 2 obtains a cost which is less than 3 C(B ) + c1 , where 2 C(B ) is the optimal cost and c1 is the cost of the largest bin if the problem satises the requirement in Eq. (10). For the sake of completeness, we list the algorithm (IFFD ) in [11] in Algorithm 3. Algorithm 3 IFFD 1: Assume that the bins are sorted in the decreasing order of their sizes. 2: Allocate all the items into bins of type 1 using the rst-t decreasing manner. Denote the resulting 1 1 1 1 sets of bins as B 1 = {B1 , B2 , , Bk1 } where Bi represents the set of items packed in the ith bin. Denote C 1 as the total cost of B 1 . Let l = 1. l 3: while l < m and max{wj : j Bkl } bl+1 do l to bins of type l + 1 4: Allocate all the items in Bkl using the rst-t decreasing manner. i 5: l = l +1; Let C l = l1 C(B i {Bki })+C(B l ) i=1 6: end while 7: let i = arg min1il C i . i1 i 8: Repack each bin in i=1 (B i Bki ) B i with the cheapest bin that can contain all items in the original bin. Let B denote the resulting solution. We next examine whether the cost functions for 802.11a and 802.11b satisfy the required conditions in Eq. (10) to help us leverage the asymptotic guarantee of 3 . Table II shows different transmission rates, their 2 corresponding SNR threshold, and the converted bin sizes and costs for 802.11a and 802.11b2 where the rate table is taken from [20], and the switching delay is assumed to be 0. From Table II, we can see that 802.11a almost satises the required conditions in [11] (except the transition from 24Mbps to 36Mbps), and 802.11b does not satisfy the conditions in general. This prevents us from directly leveraging the performance guarantee in [11] for 802.11a
Note that the SNR in the rst column is in log scale but the SNR used in the third column is in linear scale.
2

partitioning problem for multicast. Algorithm 2 shows a framework for beam multicasting based on a generalized bin-packing algorithm. While the bin-packing problem and even the variablesized bin-packing problem are well studied, the study of GCVS-BP is very limited. In the following, we rst discuss an APTAS algorithm in [7] for the GCVSBP problem, which is of very high complexity. We then show that the IFFD algorithm in [11], which was designed for a special class of GCVS-BP problems, can be applied under relaxed conditions to obtain a weaker asymptotic approximation factor of (1.5+ ). Finally, we develop a new algorithm that solves the general GCVSBP problem and achieves asymptotic 2-approximation guarantee. Epstein and Levin [7] provided an APTAS (Asymptotic Polynomial Time Approximation Scheme) for a generalized bin-packing problem, which achieves asymptotic performance ratio (1 + ) in polynomial time for any > 0. Therefore, if we apply this APTAS scheme in the framework Algorithm 2, we obtain an APTAS solution for the beam-multicast problem, as captured by the following corollary. Corollary 1: There exists an APTAS scheme for the general ASP problem (7). However, the solution in [7] requires high complexity (which is exponential in 1/, representing the tradeoff between sub-optimality and complexity). Besides the high complexity, the procedure in [7] is very involved 6

TABLE II B IN SIZES AND COSTS FOR DIFFERENT WIRELESS TECHNOLOGIES .

SNR Threshold
m (dB) dB 6.02 7.78 9.03 10.79 17.04 18.8 24.05 24.56 2.01 5.03 9.02 12.0

Rate
Rm 6 9 12 18 24 36 48 54

Bin size
bm =
1 m

Bin cost
cm =
1 Rm cm bm

802.11a 0.25 0.167 0.125 0.083 0.0182 0.0132 0.004 0.0035 802.11b 1 0.631 2 0.316 5.5 0.126 11 0.063

0.167 0.111 0.0833 0.0556 0.0417 0.0278 0.021 0.0185 1 0.5 0.1818 0.0909

0.667 0.667 0.667 0.667 2.290 2.11 5.29 5.29 1.5849 1.5811 1.4442 1.4408

and 802.11b. However, we prove the following relaxed results for the IFFD algorithm, which can then be applied to both 802.11a and 802.11b according to Table II to obtain performance guarantees. Theorem 3: If there exists > 0 such that (1 + )ci /bi c1 /b1 for all i > 1, then C(BIF F D ) < 3 ( 2 + )C(B ) + c1 , where = 3 . 2 The proof is given in Appendix B. Comments: There are two key relaxations in Theorem 3 compared to the relevant Theorem 6 in [11], which warrants a new proof. First, it is not required that the ci largest bin has the lowest cost-to-size ratio (i.e., bi ). It is sufcient if the cost-to-size ratio of smaller bins is 1 within a fraction (i.e., 1+ ) of that of the largest bin. Second, we do not require that the cost-to-size ratio is non-decreasing as bin sizes decrease. We only need to compare the cost-to-size ratio of other bins with the largest bin. Therefore, if we set = 0, we obtain the same conclusion as Theorem 6 in [11] but with more relaxed assumptions. Now using Theorem 3, and referring back to Table II, we nd that the minimum values are 0.1 and 0 for 802.11a and 802.11b, respectively. Hence, the asymptotic approximation factors for the two wireless systems are 1.65 and 1.5, respectively, which is summarized in the following Corollary. Corollary 2: For the beam partitioning problem under ASP (7) with rate function dened according to 802.11a/b, if zero switching delay is assumed, Algorithm 2 using IFFD as a GCVS-BP solver has asymptotic approximation guarantee of 1.65 and 1.5, respectively. An Asymptotic 2-approximation Algorithm: IFFD pro7

vides an efcient multicasting solution, whose performance guarantee depends on the specic technology (through ). However, a variety of technologies are forthcoming such as 802.16m, LTE, LTE-advanced, and 60GHz systems. While it is possible to obtain a for each future system, it is more desirable to develop algorithms that have a uniform guarantee regardless of the cost-to-size ratio constraint and hence the underlying technology. We next devise such a novel algorithm for the GCVS-BP problem with a uniform guarantee. The algorithm, named MEBC, is illustrated in Algorithm 4, and proceeds as follows. Without loss of generality, we assume that those bins with both smaller or equal sizes and larger costs than any other bin are excluded from consideration and that the bins are sorted in decreasing order of their sizes. First, we nd the set of most efcient bin types T (Step 1). A bin type is most efcient if it has the smallest cost-to-size ratio among all bin types with equal or larger sizes. Therefore, multiple most efcient bins exist. Formally, we dene T = {k1 = 1, k2 , , kl } such that, for each bin type ck ci kj , bkj < bi for all i < kj . j We then pack the items in the order of decreasing sizes (Step 2). When packing each item, we nd the rst available bin that is not empty and can hold the current item (Step 3). If no such non-empty bin can be found, a new bin is created to hold the item. When creating a new bin, we create the most efcient bin with the smallest size that is just enough to hold the current item (Step 4). Finally, we repack every bin using a bin (not limited in T ) that has the smallest cost but can hold all items in the corresponding original bin (Step 5). Algorithm 4 MEBC (Most Efcient Bin Created) 1: Compute the set T of most efcient bin types. 2: Pack the items in the order of decreasing sizes. 3: For each item with size sm , nd the rst bin that is not empty and has sufcient space to hold it. 4: If no existing bin can hold the item, create a bin of type kj T such that bkj sm > bkj+1 and put the item in. 5: Re-pack every bin using a bin that has the smallest cost and can hold all objects in the corresponding original bin. We next show that Algorithm 4 achieves an asymptotic 2-approximation. Note that this easy-to-implement, asymptotic 2-approximation algorithm, which does not impose any constraint on the bin size and cost relation-

ship, may be of independent interest in the area of binpacking problems and their applications. Theorem 4: Let C(B ) be the optimal cost of bin packing and C(BMEBC ) be the cost using the MEBC algorithm. We have
C(BMEBC ) < 2C(B ) + c1

The proof is shown in Appendix C. D. Continuous Rate Function We now consider the multicasting problem (7) under ASP for a continuous rate function R(), where we assume that the rate function R() is continuous and non-decreasing. We develop sufcient conditions under which (i) one group, (ii) K groups (where K is the number of beams), or (iii) any number of groups, forms the optimal partitioning scheme. Theorem 5: Assume that R() 0 is non-decreasing L over [0, ). Dene g() = (W + R() ). If g() is an increasing function of SNR value , the optimal partition has one group including all single-lobe beams. If g() is a decreasing function of SNR value , the optimal partition has K groups and each group contains one single-lobe beam. If g() is a constant, any partition is optimal. Before we prove the theorem, we show some special examples of the rate functions. The following corollary shows that for a wide class of continuous rate functions, the optimal partition is to have only one beam group. Corollary 3: If the rate function R() 0 is nondecreasing and concave over [0, ), it is optimal to have only one group. Special examples that fall in this category include the Shannon channel capacity, R() = B log2 (1 + ), and the modied Shannon capacity, R() = B log2 (1 + /) where > 1 represents the gap between the channel capacity and the actual coding. If W = 0 and R() = C (which can be viewed as the approximate Shannon channel capacity at the low SNR regime), any partition is optimal. Proof: Assume R() 0 is concave over [0, ). In order to show that the optimal partition is to have one beam group, it is sufcient to show that g () 0. Since
g () = W + L R() R () , R2 ()

If W = 0 and R() = C , we obtain that g() = L/C is a constant, so any partition is optimal. Proof of Theorem 5: For continuous rate functions, Problem (7) can be viewed as a continuous version of GCVS-BP problem, in which there are innitely many types of bins and the bin sizes can take any continuous positive real value. Now for any SNR and rate function R(), it corresponds to a bin size s = 1/ and cost W + L/R(). As R() is a non-decreasing function of , indicating the bin cost is non-decreasing as the bin size increases, in the optimal solution, the bin size should be equal to the total size of all items in the bin. As there is no extra space left in a bin, it is optimal to choose bins that have the lowest cost-to-size ratio (i.e., (W + L/R(1/s))/s). Therefore, if the cost-to-size ratio of bins is a nonincreasing function of bin size s, the best choice is to have only one bin containing all items, with a size that is equal to the total size of all items. If the cost-to-size ratio of bins is an non-decreasing function of s, the optimal packing is to choose as small bin sizes as possible. Thus, the optimal solution in this case is to put one item in each bin, with a size equal to the size of the item. If the costto-size ratio of bins is a constant c, it does not matter which bins are chosen. Thus the total cost is always given by the total size of all items to be packed multiplied by c, regardless of how they are packed. Finally, notice that the cost per unit size of bin, (W + L/R(1/s))/s, is increasing (decreasing, or constant) with respect to s if and only if (W + L/R()) is decreasing (increasing, or constant) with respect to . This completes the proof. V. T RACE - DRIVEN P ERFORMANCE E VALUATION In this section, we evaluate the performance of the proposed algorithms through trace-driven simulations where the signal SNR trace is obtained from an experimental beamforming system. The experimental testbed consists of an eight-element Phocus Array [6] from Fidelity Comtech as the access point (AP) and multiple laptops with omni-directional antennas as mobile clients. The Phocus Array contains 8 antenna elements and is capable of providing eight 45-degree single-lobe beam patterns that are approximately non-overlapping. Figure 1 shows the testbed in an outdoor parking lot as well as the locations of both the AP and clients used in the experiments. We compare our algorithms with the GREPd and GRASP2 algorithms in [20] for the EQP model and ASP 8

it is sufcient to show that R() R () 0. By the convexity theory (e.g., Proposition B.3 in [2]),
0 R(0) R() + (0 )R ().

(11)

Therefore, the sufcient condition for having one beam group is satised.

802.11a 5 4.5 4 Multicast time (ms) Multicast time (ms) 3.5 3 2.5 2 1.5 DPEQP IFFDASP MEBCASP GREPd GRASP2 10 20 30 Number of users 40 10 14

802.11b

12

6 DPEQP IFFDASP MEBCASP GREPd GRASP2 10 20 30 Number of users 40

Fig. 1.

Outdoor testbed and AP/client locations.

1 0

2 0

Fig. 2.

Uniform user distribution in all beam sectors.


802.11a 802.11b 9 8 7 6 5 4 3 2 2 DPEQP IFFDASP MEBCASP GREPd GRASP2 4 6 Number of beam sectors (K) 8

model, respectively. We do not include the other wellknown beamforming multicasting algorithm beamcast [17] as it was shown in [20] that GREPd and GRASP2 outperform beamcast. We carry out simulations based on the SNR measurements at all receivers locations in the testbed but randomly pick a subset of users for each simulation run. We simulate both the continuous and the discrete rate cases, but only report the results for the discrete case as almost all practical wireless systems use discrete rate tables. In the resulting gures, all delays are calculated based on transmitting a 1500-byte multicast message. All results are averaged over ten simulation runs. A. Performance without Switching Delay In this section, we compare the performance in the ideal situation where there is no delay when switching beam patterns. Uniform node distribution: We rst consider the case where the users are uniformly distributed across all 8 beam sectors. Figure 2 shows the aggregate delay for 802.11a and 802.11b systems. It can be seen that our DP-EQP algorithm consistently outperforms the greedy algorithm GREPd, with a delay reduction of about 10% for both 802.11a and 802.11b. For the ASP algorithms, it is interesting to observe that both IFFD-ASP and MEBCASP obtain signicant improvement over GRASP2. The average improvement of both IFFD-ASP and MEBCASP, compared to GRASP2, is over 15% and 20% for 802.11a and 802.11b systems, respectively. From Fig. 2, we can also see that IFFD-ASP and MEBC-ASP perform similarly in most scenarios and MEBC-ASP slightly outperforms IFFD-ASP when the number of clients increases. Moreover, we can see that when the number of clients increases, the improvement of IFFD-ASP and MEBC-ASP over other schemes increases. This is not surprising as when the number 9

3.5 Multicast time (ms) Multicast time (ms) DPEQP IFFDASP MEBCASP GREPd GRASP2 4 6 Number of beam sectors (K) 8

2.5

1.5

1 2

Fig. 3. Clustered user distributions where users are clustered in K beam sectors.

of clients increases, the number of possible partitions increases and so does the potential for improvement. Clustering node distribution: We also investigate the case where mobile clients locations are clustered. To model user clustering, we randomly draw 30 users from a randomly selected K beam sectors where K 8. Figure 3 compares the aggregate delay of transmitting a 1500byte message. Similar patterns are observed as in the case of uniform node distributions. It can also be seen that when the system is less clustered (i.e., larger number K of beam sectors), the improvement of IFFD-ASP and MEBC-ASP (compared to other schemes) increases. Therefore, we conjecture that the proposed algorithms are more benecial to systems with larger number of beams such as 60GHz systems [9], [23]. B. Performance with Switching Delay We next show how the switching delay affects the performance of all algorithms. Figure 4 plots the multicast transmission time vs. switching delay for different algorithms. It can be seen that the multicast transmission

802.11a 9 8 7 6 5 4 3 2 0 DPEQP IFFDASP MEBCASP GREPd GRASP2 16 15 14 Multicast time (ms) 13 12 11 10 9 8 7 500 Switch delay (us) 1000 6 0

802.11b DPEQP IFFDASP MEBCASP GREPd GRASP2


7 6 Packet loss (%) 5 4 3 2 1 0 0

802.11a IFFDASP MEBCASP DPEQP 5 4.5 4 Packet loss (%) 3.5 3 2.5 2 1.5 1 10 20 30 Number of users 40 0.5 0

802.11b IFFDASP MEBCASP DPEQP

Multicast time (ms)

500 Switch delay (us)

1000

10 20 30 Number of users

40

Fig. 5. Fig. 4. Comparison of multicast time with switching delay.

Impact of channel state uncertainty.

VI. E XPERIMENTAL E VALUATION time increases almost linearly as the switching delay increases for all algorithms. As the switching delay increases, IFFD-ASP shows 20% and 25% improvement over GRASP2 for 802.11a and 802.11b consistently. While IFFD-ASP and MEBC-ASP achieve nearly the same delay in 802.11a systems, IFFD-ASP obtains much smaller delay than MEBC-ASP in 802.11b systems with a large switching delay. A little investigation shows that the performance gain is due to Steps 3-6 in IFFD-ASP, which attempt to break the last bin into multiple smaller bins in anticipation that it may not be sufciently lled. However, we note that these steps can also be applied to MEBC-ASP. In this section, we conduct small-scale experiments to evaluate the developed algorithms. A. Experimental setup Testbed: We perform experiments using a testbed consisting of an 802.11b/g access point with a beamforming antenna from Fidelity-Comtech [6] and three clients with D-Link DWL-AG660 802.11a/b/g cards in the outdoor environment shown in Figure 1. Both the AP and clients run Ubuntu 8.04 and Madwi WLAN drivers. We use iperf as the trafc generator and the athstats Madwi utility to obtain packet statistics (e.g., packet loss rates). Phocus Array Antenna: An eight-element Phocus Array Antenna provided by Fidelity Comtech [6] is employed as the AP, as shown in Figure 6. The magnitude and phase of signal on each element can be set separately to form a special beam pattern. The antennas rmware provides multiple pre-setting switched beamforming patterns with various direction, power and width. Figure 7 shows one of these patterns. Table III is the conguration of this pattern, where Mag represents the percentage of maximum transmit power used on each element. Using the SDK tools provided with the antenna, we can add customized patterns to the antenna through a command line interface. Measurement: We use RSSI(Receive Signal Strength Indication) to represent the SNR values in our experiments. Latest Madwi driver provides this measurement by subtracting the measured noise level from the signal strength (both in dBm). Packet loss rate is computed as (1 trecv ), where tsent is the number of packets tsent sent by the antenna and trecv is the number of packets successfully received at a client. Both numbers are taken from the athstats Madwi utility to obtain an accurate 10

C. Uncertainty of Channel Conditions In real (especially mobile) environments, wireless channel conditions vary and cannot be estimated precisely. In this set of simulations, we generate the wireless channel SNR (in dB) according to a Gaussian distribution based on the measured mean and variance values at each client, but the algorithms only use the mean value of the SNR. To emulate the packet transmission errors, we record a packet loss if the randomly generated SNR of a client is smaller than the SNR threshold of the MCS computed by each partitioning algorithm. Figure 5 shows the packet loss rate of our algorithms for the uniformly distributed user distributions. It can be observed that the IFFD-ASP is most robust among the three algorithms evaluated. This is because, IFFD-ASP always tries to use the largest bin to pack items in the rst step, which corresponds to using the lowest MCS to transmit packets. Thus, it is the most reliable under varying (fading) wireless channel conditions.

TABLE III PATTERN CONFIGURATION ON ANTENNA ELEMENTS .

Element
1 2 3 4 5 6 7 8

Mag
98 51 34 41 41 35 54 100

Phase(Deg.)
46 86 -82 -50 -50 -78 85 43

TABLE IV R ATE M AP Fig. 6. Phocus Array Antenna with eight elements

RSSI Threshold
(dB) 25.5 27.33 28.67 29.67 31.17 34.00 38.33 40.33

Rate
(Mbps) 6 9 12 18 24 36 48 54

Fig. 7.

Pattern Example

measure at the PHY layer. A RSSI-based rate table is obtained through measurements and is determined by nding the minimum RSSI value for each rate such that the packet error rate is less than 10%. Table IV shows the rate map we get. B. Experiment Steps After obtaining the RateMap as in Table IV, we execute the following steps in our experiments. Step 1. In the outdoor environment shown in Figure 1, we deploy the AP (with beamforming antennas) and three clients around the AP but well spaced between each other. Step 2. Every client measures the RSSI values under each of the eight single-lobe beam patterns as well as the omni-directional pattern and reports the highest RSSI value and the best beam index to the AP. Step 3. The AP applies DP-EQP, IFFD-ASP, MEBCASP and Optimal-ASP with the reported RSSI values and obtains the multicast delay and beam partitioning for every algorithm, where the Optimal-ASP algorithm enumerates all possible ways of partitioning and nds the best one. A simple broadcast scheme with the omnidirectional beam pattern is also run for comparison. 11

Step 4. A le containing about 1000 packets is multicasted from the AP to the clients using the partitions obtained in Step 3. Average packet delivery ratios are measured at each client to verify the correctness of the beam partitioning algorithm. The multicast is performed by sending packets to a mobile placed very close to the AP (so that very few re-transmissions happen) and conguring all three intended clients in the promiscuous mode to overhear the packets. C. Performance Metrics We evaluate our beamforming multicast algorithms with the following three metrics. 1. Multicast Delay is the average time of transmitting a packet with B bytes with all beam partitions (computed by each algorithm). Specically, it is dened as
G g=1

B , Rg

(12)

where G is the number of partitions and Rg is the transmission rate assigned to partition g. Here we omit the switching delay which is typically 100-150us [20]. B is set to 1500 x 8 bits in our experiments. 2. Average Delivery Ratio (ADR) is the ratio of packets that are received successfully by the clients. In a multicast session, we transmit a le with

Multicast Delay 4

Delivery Ratio 3.5

Number of Partitions 3 OptimalASP IFFDASP MEBCASP DPEQP

Packets Delivery Ratio(%)

Number of Partitions

Multicast Delay(ms)

Omni OptimalASP IFFDASP MEBCASP DPEQP

100

98

Omni OptimalASP IFFDASP MEBCASP DPEQP

2.5 2 1.5 1 0.5

96

94 4 6 8 10 Transmit power(dBm)

4 6 8 10 Transmit power(dBm)

4 6 8 10 Transmit power(dBm)

(a)
Fig. 8.

(b)
Experiment results

(c)

T packets. Let Ti be the number of packets successfully received by client i. The ADR is then , (13) 3T 3. Number of partitions may have considerable impact when switching delay of beam patterns is taken into account. In that case, the more partitions in the output of an algorithm, the longer time the multicast procedure takes. ADR =
3 i=1 Ti

the omni-directional multicast scheme. It is interesting to observe that while the optimal solution achieves less delay at 3dBm transmit power than the other ASP algorithms, it also has lower delivery ratio because more aggressive MCSs are used in the optimal solution. The results for the omni-pattern algorithm at 3dBm transmit power are not shown because not all clients are covered by the omni-pattern at this power level. Finally, the number of partitions of omni-broadcast is always 1 and not plotted in the gures. VII. R ELATED W ORK Wireless Multicasting in Communication Theory: In the communication area, many researchers have studied the problem of multicast/broadcast with adaptive/beamforming antennas [22], [18], [19], [24], [14]. Most of the works assumed adaptive beamforming or MIMO techniques, which require higher complexity to compute the optimal beam or antenna weights and need detailed complex-value channel feedback, as opposed to switched beamforming, which is the focus of this work. Link layer algorithms for wireless multicasting: Many works have focused on link-layer algorithms for enhancing wireless multicasting [16], [5], [10], [4], [13], [3]. Park et al. proposed a new rate-adaptation algorithm for multicasting multimedia content. The works [5], [10] developed efcient feedback mechanism to improve the multicasting reliability. Chaporkar et al. [4] designed a scheduling policy for multicasting in ad hoc wireless networks. Li et al [13] presented efcient resource allocation algorithms for multicasting scalable video streams. Chandra et al. [3] built a WiFi prototype implementation of wireless multicasting and also solved 12

D. Evaluation Results As it is difcult to eliminate all processing delay, contention delay, and MAC-layer re-transmission backoff delay, the delay values reported in Fig. 8(a) are those generated by all algorithms. Nevertheless, the packet delivery ratios in Figure 8(b) are obtained from the measurements at all clients. We evaluate the performance of ve multicast algorithms: simple omni-broadcast, IFFD-ASP, MEBC-ASP, DP-EQP, and optimal-ASP. We conduct experiments with transmit power ranging from 3 dBm to 11 dBm. For each experiment, a le containing about 1000 packets each with 1500 bytes is multicasted to three clients. Figures 8 (a), (b), (c) show the multicast delay, average delivery ratio, and number of partitions, respectively. It can be seen from Fig. 8 that, i) all algorithms achieve the target delivery ratio 90% which is used for constructing the rate-table, ii) both IFFD-ASP and MEBC-ASP algorithms achieve similar delay as the optimal algorithm (except for the 3dBm transmit power), and iii) the delay of all ASP algorithms is smaller than that of DP-EQP which is in turn smaller than that of

several practical problems including the AP association problem. Although all these works have their own merits, they consider only omni-directional antennas. Beamforming multicasting algorithms: Only a few recent works looked at the integrated problem of beamforming and multicasting [8], [17], [20]. While Hou et al. [8] developed new multicasting routing algorithms exploiting beamforming antennas, the works [17], [20] aimed to design link-layer multicast scheduling algorithms with switched beamforming antennas. Sen et al. [17] presented a rst-cut solution, by performing a omnidirectional transmission followed by one or a few singlelobe sequential directional transmissions to cover the clients left behind from the initial omni-transmission. Sundaresan et al. [20] provided a rigorous formulation of the switched beamforming multicasting problem and proposed several algorithms to solve the problem, under different models. In this work, we adopt the problem formulation in [20] but make several signicant contributions, including a much more efcient optimal algorithm under the EQP model and several asymptotic approximation solutions under the ASP model. VIII. C ONCLUSION We have studied the problem of multicasting with beamforming antennas in wireless networks. We consider both the EQP model and ASP model. Under the EQP model, we obtain optimal algorithms based on dynamic programming for arbitrary rate functions. Under the ASP model, we prove that the general problem is NP-hard and obtain approximation solutions for discrete rate functions. For the continuous rate function under ASP model, we also develop a set of sufcient conditions under which the optimal solution has (i) 1 group, (ii) K group (where K is the number of beams), and (iii) arbitrary number of groups. In particular, we show that if the rate function is continuous, non-decreasing and concave, it is optimal to have only one group. The effectiveness of our algorithms is evaluated through both experiments and trace-driven simulations, and signicant improvement is observed over recently proposed algorithms. R EFERENCES
[1] J. Akkermans and M. Herben. Planar beam-forming array for broadband communication in the 60 ghz band. In European Conf. Antennas and Propagation (EuCAP07), Nov. 2007. [2] D. Bertsekas. Nonlinear Programming (2nd Edition). Athena Scientic, April, 2004. [3] R. Chandra, S. Karanth, T. Moscibroda, V. Navda, J. Padhye, R. Ramjee, and L. Ravindranath. Dircast: A practical and efcient wi- multicast system. In IEEE ICNP, Princeton, 2009.

[4] P. Chaporkar, A. Bhat, and S. Sarkar. An adaptive strategy for maximizing throughput in mac layer wireless multicast. In ACM Mobihoc, 2004. [5] A. Chen, D. Lee, G. Chandrasekaran, and P. Sinha. Himac: High throughput mac layer multicasting in wireless networks. In IEEE MASS, Oct. 2006. [6] F. Comtech. Phocus array system. In http://delitycomtech.com. [7] L. Epstein and A. Levin. An aptas for generalized cost variable sized bin packing. SIAM Journal on Computing, 38:411428, 2008. [8] Y. Hou, Y. Shi, H. D. Sherali, and J. E. Wieselthier. Multicast communications in ad hoc networks using directional antennas: A lifetime-centric approach. IEEE Transactions on Vehicular Technology, 56(3), May 2007. [9] IEEE-802.11. Ieee very high throughput 60ghz (vht60). [10] S. Jain and S. R. Das. Mac layer multicast in wireless multihop networks. In IEEE COMSWARE, Jan. 2006. [11] J. Kang and S. Park. Algorithms for the variable sized bin packing problem. European Journal of Operational Research, 147(2):365372, June 2003. [12] S. Lakshmanan, K. Sundaresan, S. Rangarajan, and R. Sivakumar. Practical beamforming based on rssi measurements using off-the-shelf wireless clients. In the 9th ACM SIGCOMM conference on Internet measurement conference, 2009. [13] P. Li, H. Zhang, B. Zhao, and S. Rangarajan. Scalable video multicast in multi-carrier wireless data systems. In IEEE ICNP, 2009. [14] E. Matskani, N. Sidiropoulos, and L. Tassiulas. On multicast beamforming and admission control for umts-lte. In IEEE ICASSP, 2008. [15] R. Motwani. Lecture notes on approximation algorithms: Volume I. Technical Report: CS-TR-92-1435, Standford University, 1993. [16] Y. Park, Y. Seok, N. Choi, Y. Choi, and J. Bonnin. Rate-adaptive multimedia multicasting over ieee 802.11 wireless lans. In IEEE CCNC, 2006. [17] S. Sen, J. Xiong, R. Ghosh, and R. Choudhury. Link layer multicasting with smart antennas: No client left behind. In IEEE International Conference on Network Protocols (ICNP), 2008., pages 5362, Oct. 2008. [18] M. Sharif and B. Hassibi. On the capacity of mimo broadcast channels with partial side information. IEEE Transactions on Information Theory, 51(2), Feb. 2005. [19] Y. Silva and A. Klein. Adaptive antenna techniques applied to multicast services in wireless networks. Frequenz, 59, 2005. [20] K. Sundaresan, K. Ramachandran, and S. Rangarajan. Optimal beam scheduling for multicasting in wireless networks. In ACM Mobicom, 2009. [21] V. V. Vazirani. Approximation Algorithms. Springer, 2001. [22] S. Vishwanath, N. Jindal, and A. Goldsmith. Duality, achievable rates, and sum-rate capacity of gaussian mimo broadcast channels. IEEE Transactions on Information Theory, 49(10), Oct. 2003. [23] WirelessHD. Wireless hd specication. Version 1.0, Oct. 2007. [24] T. Yoo and A. Goldsmith. On the optimality of multiantenna broadcast scheduling using zero-forcing beamforming. IEEE Journal on Selected Areas in Communications, 24(3), March 2006. [25] H. Zhang, Y. Jiang, S. Rangarajan, and B. Zhao. Multicast video delivery with switched beamforming antennas in indoor wireless networks. In Submitted to IEEE Infocom 2011.

13

A PPENDIX A A PPROXIMATION
ALGORITHMS

We dene some terminologies for approximation algorithms used in this paper, largely following the denitions in [15]. We consider the problem of minimization of an non-negative objective function. For a given problem instance I , we use A(I) and OPT(I) to denote the objective values of the solution of an algorithm A and the optimal solution, respectively. If for r > 1 and any instance I ,
A(I) r OPT(I),

t(Bj ), the minimum cost of holding these objects in the optimal solution is t(Bj ) . 1+ Proof. The minimum cost per unit bin size for all types of bins is 1 c1 1 ci = . min bi 1 + b1 1+

Therefore, the minimum cost for a total object size of t(Bj ) is t(Bj ) . 1+ We prove the theorem by conditioning on different cases. Denote the optimal cost is C(B ). If for every bin j except the last bin k1 , t(Bj ) 2 c(Bj ), then 3
k1

we call the algorithm A is an r -approximation algorithm, or A has performance ratio r . Let


r = inf{s > 1 : N0 such that for all I with OPT(I) N0 , and A(I) s OPT(I)}.

c(Bj ) 1 +
j=1

3 2

k1 1 j=1

3 t(Bj ) 1 + (1 + )C(B ) 2

We call the algorithm A is an asymptotic r approximation algorithm, or A has asymptotic performance ratio r . As a special case, if
A(I) r OPT(I) + C,

where C is a constant, the algorithm A has asymptotic performance ratio r . An asymptotic approximation scheme is an algorithm A that takes as input both the instance I and an error bound > 0, and has asymptotic performance ratio (1 + ). An APTAS (Asymptotic Polynomial Time Approximation Scheme) is an asymptotic approximation scheme {A } where each algorithm A has asymptotic performance ratio 1 + and runs in time polynomial in the length of the input instance I . A PPENDIX B P ROOF OF T HEOREM 3 Without loss of generality, we assume that b1 = c1 = 1 (Otherwise, we can always normalize all bin sizes with respect to b1 and all cost with respect to c1 ). Now we only need to prove that C(BIF F D ) (1+) 3 C(B )+1. 2 As the steps in iterations 3-6 do not increase the cost, it is sufcient to prove the theorem without them and simply let i = 1 in step 7. As we only consider the bins in B 1 , we omit the superscript in the following discussion to ease the notations. For each allocated bin Bj , we use c(Bj ), b(Bj ), t(Bj ) to represent the cost, size, and the total sizes of objects in the bin, respectively. First we show a lemma. Lemma 2: Assume that the conditions in Theorem 3 are satised. If the total size of the objects in a bin is 14

where the last inequality follows from Lemma 2. The conclusion holds. Therefore, we only need to consider the case that there exists some k < k1 such that t(Bk ) < 2 c(Bk ). Without 3 loss of generality, we assume k is the largest index such that the condition satises (but still k < k1 ). We claim that t(Bk ) > 1/2. Because otherwise, any item in the bin Bk+1 should have been put into Bk . We also show that Bk only contains one item. Otherwise, Bk must contain an item with size less than 1/3, since the total sizes of 2 the objects in Bk is t(Bk ) < 2 c(Bk ) 3 . This indicates 3 that for all j < k, its total object size is at least 2/3 (otherwise, the second-largest item in Bk should have been put into one of these Bj s). Further, if we put all objects in the last bin (with index k1 ) to the kth bin, the total object sizes in the kth bin will exceed 1 (otherwise, the bin k1 will not be created). So it can be viewed that every bin j except the last bin has total object size 2 t(Bj ) 2/3 3 c(Bj ). Therefore,
k1

C(BIFFD ) = 1+ 3 2
j=1 k1 j=1

c(Bj ) 3 t(Bj ) 1 + (1 + )C(B ). 2

Now we conclude that Bk contains only one item and 1/2 t(Bk ) 2/3c(Bk ). So the cost of Bk is at least 3/4. Since Bk is obtained after step 8 of IFFD, for all bins with size at least t(Bk ), their cost is at least 3/4 (Otherwise, they should have been used to hold the objects in Bk with lower cost). It is also clear that all bins with index less than k contain one item with size larger than t(Bk ) due to the IFFD algorithm. Therefore, in order to contain the one big item in each bin j k,

at least one bin with cost 3/4 is required, and any two of them cannot be put into one bin because their sizes is larger than 1/2. As a result, the minimum cost of holding 3 these big items in the optimal solution is at least 4 k. Further, all of the items in bin j > k have sizes larger than 1t(Bk ) (Otherwise, they should have been packed to the bin k). Therefore, none of these items can be put in the bins holding the previous big items from bins j k Thus, the minimum cost of these items in bin P
j > k is
k1 j=k+1

t(Bk ) . 1+

As a result, the optimal cost is

k1 3 j=k+1 t(Bk ) . C(B ) k + 4 1+ The resulting cost from the algorithm IFFD is k1 1

(14)

C(BIF F D ) k + k+ 3 2

c(Bj ) + 1
j=k+1 k1 1

t(Bj ) + 1
j=k+1 k1

3 3 j=k+1 t(Bj ) < (1 + )( ( k + )) + 1 2 4 1+ 3 (1 + ) C(B ) + 1 (15) 2 where the second inequality is because for all k < j < k1 , t(Bj ) 2 c(Bj ), and the last inequality is from Eq. 3 (15). This concludes the proof.

assume that the most efcient set of bin types T (see the descripton of Algorithm MEBC) is the set of all bin types, that the bins are sorted in the decreasing order of both sizes and cost-to-size ratios, and that every bin in T is used at least once. Consider the packing result obtained with MEBC. Recall that bj and cj (j = 1, , k) denote the size and cost of the j th type of bin in T . Let hj denote the set of items whose sizes are in (bj+1 , bj ]. The items in hj can fall in three categories: xj , yj and zj . Items in xj are put into the bins of type 1, , j 1. These items are used to ll up the holes in the previously generated bins. Items in yj are put into bins with type j but are not in the last bin with type j . Items in zj are put into the last bin with type j . Note that we also use xj , yj , and zj to denote the total sizes of all items in them. It is not hard to observe that i) bins containing yj are all less than half empty, and ii) zj1 + xj + zj > bj1 (otherwise, all items in xj , zj can be put into the last bin of type j 1 with MEBC). Since all items in zj1 +xj +zj are larger than bj+1 , their cost-to-size ratios are lower bounded by cj /bj . Therefore, C (zj1 +xj +zj ) bj1 cj /bj > cj . Summing up for all j = 1, , k, noting that C (x) 1 2 C (2x), and applying Lemma 3, we have
k

C (
j=1

(zj + xj ))

1 1 C ( (zj1 + xj + zj )) > 2 2 j=2

cj (16)
j=2

A PPENDIX C P ROOF OF T HEOREM 4 We use variables x, y, z to denote both the sets of items and the total sizes of the items in the sets and use (+) sign to represent both the union of sets of items and the summation of their sizes. Denote C (x) as the optimal cost of packing items in a set x. Since an item can only be put into bins with larger or equal sizes, its optimal cost-to-size ratio is lower bounded by the most efcient bin that can contain it and has the smallest size. We rst prove a technical lemma. Lemma 3: Let x and y be two sets of items and all items in sets x and y has a minimum cost-to-size ratio of r1 and r2 , respectively. Then the optimal cost C (x + y) of packing the items in the two sets x + y is at least r1 x + r2 y . The proof of the lemma is straightforward and is omitted. Note that the lemma can be easily generalized to the case of multiple sets. We now proceed to prove the theorem. Our proof does not consider Step 5 of the algorithm, which does not increase the total cost. Without loss of generality, we 15

Let nj be the number of bins of type j used by MEBC. The number of bins containing items yj is then nj 1. Since type j is the most efcient bin with the smallest size for all items in yj and all bins containing yj are less than half-empty, we get
C (yj ) yj cj bj cj cj (nj 1) (nj 1) . bj 2 bj 2

Applying Lemma 3 and combining with Eq. (16), we have C ( k yj ) k (nj 1)cj /2 and j=1 j=1
k

C(B ) = C (
j=1 k

(xj + yj + zj )) cj + 2
k

j=2

(nj 1)
j=1

cj 2 cj . 2

= (n1 1)

c1 + 2

nj
j=2

Finally, note that the cost of the MEBC is just


k

C(BMEBC ) =
j=1

nj cj 2C(B ) + c1 .

This completes the proof.

You might also like