Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

2018 IEEE 19th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)

User Scheduling in Massive MIMO


Hong Yang
Bell Laboratories, Nokia, 600 Mountain Avenue, Murray Hill, NJ 07974 USA
h.yang@nokia-bell-labs.com

Abstract—Massive MIMO relies on nearly orthogonal user interval [11], [12]. A solution to this problem is to serve users
channels to achieve unprecedented spectral efficiency. But in with similar channels separately in different coherence time
LoS (line-of-sight) environment, some users can be subjected slots. To successfully realize this solution one needs:
to similar channel vectors. Serving users with similar channel
vectors simultaneously can severely compromise the throughput • An effective and simple mechanism to quickly identify
performance to all users. We propose a scheduler that identifies users with similar channels;
users with similar channels and serves them in separate time slots • An effective and simple scheduler to schedule users with
with properly assigned data rates, while aiming to provide fair similar channels to serve in separate coherence slots, and
service to all users and maximize the system spectral efficiency
at the same time. Simulation results show the effectiveness of with proper rate assignment to each user.
the scheduler on both downlink and uplink of a single cell The mechanisms for these two tasks must be simple so it
Massive MIMO with MR (maximum ratio) processing or ZF can be done fast, within the coherence time of the system,
(zero-forcing) processing, and that channel correlation threshold which is typically in the order of milliseconds for mobile
for scheduling users is an important design parameter that can
be fine-tuned to optimize the user throughput performance. networks. Additionally, the scheduler should maximize the
Index Terms—Massive MIMO, scheduler, line-of-sight, power spectral efficiency as much as possible. An algorithm was
control. proposed in [11] for determining which users should be
dropped from service so that the user throughput would not
I. I NTRODUCTION be compromised for the remaining users. But how do we
provide a fair service to all users and maximize the total
Scheduler design, both in physical and application layers, spectral efficiency? We propose an answer to this question
has been an important research area for the past and current in this paper. The proposed scheduler does not depend on
generations of multi-user wireless technologies. For exam- the precoding/decoding schemes, and is applicable to both
ple, for multi-user MIMO (multiple-input/multiple-output), [1] downlink and uplink. We evaluate the performance of the
considered downlink scheduling with equal power for the scheduler on MR (maximum ratio) and ZF (zero-forcing)
scheduled users; [2] extended proportional fair scheduling precoding/decoding schemes via extensive simulations.
for the uplink operation; [3] studied the scheduling problem,
taking into account the requirements from both physical and II. M ASSIVE MIMO IN L O S
application layers; [4] derived capacity achieving scheduling We assume a single-cell Massive MIMO with an M -
algorithms for zero-forcing beamforming as the number of antenna array base station which serves K simultaneous and
users tended to infinity; [5] investigated user scheduling to autonomous users.
mitigate pilot contamination in Massive MIMO [6] [7]; [8]
explored the benefits of multi-user precoding for millimeter A. Effective SINR
wave MIMO systems. However, so far there has been little Let
research in the physical layer scheduler design for Massive G = (g1 · · · gK ) ∈ CM ×K
MIMO, a core wireless technology for 5G. We believe that this
disparity is mainly due to: (1) with precision beamforming of be the M ×K channel matrix between the M -antenna array at
Massive MIMO, max-min power control can be used to deliver the base station and the autonomous K users, where gk ∈ CM
equal throughput to all users [9]; (2) traditional proportional is the channel vector between the M -antenna array at the base
fair scheduler, which takes advantage of small-scale channel station and the kth user, k = 1, · · · , K.
variation to schedule users during up fade, provides little Effective SINR for MR processing and ZF processing have
benefit in increasing system spectral efficiency due to channel been derived in [11], [13] and are summarized in Table I.
hardening [10] in Massive MIMO. In Table I, [A]k,k denotes the kth diagonal element of a
But there are situations where a physical layer scheduler is matrix A and superscript H denotes conjugate transpose. Let
necessary for Massive MIMO. For example, in predominantly R0+ denotes the set of all nonnegative real numbers.
LoS (line-of-sight) environment, with non-negligible probabil- η = (η1 , · · · , ηK )T ∈ RK
0+
ity some users can experience similar channel fading [11]. In
such a case, the channel matrix becomes nearly singular. The is the downlink power control vector with total power con-
result is a severe degradation in fairness and/or system spectral straint
efficiency if all the users are served in the same coherence η1 ≤ 1,

978-1-5386-3512-4/18/$31.00 ©2018 IEEE

Authorized licensed use limited to: National Tsing Hua Univ.. Downloaded on August 12,2021 at 09:53:07 UTC from IEEE Xplore. Restrictions apply.
2018 IEEE 19th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)

TABLE I has a solution η̆ ∈ RK


0+ that satisfies η̆∞ ≤ 1. The uplink
E FFECTIVE SINR power control is given by η̆. Here
Maximum Ratio Zero-Forcing
Λ̆ = diag(g1 −2 −2
2 , · · · , gK 2 )Λdiag(g1 2 , · · · , gK 2 ),
2 2
gk 22 ρd ηk ρd η k
Downlink T
 |gH gl |2 K
 b̆1 = s̆1 g1 −2 −2 −2
2 , s̆2 g2 2 , · · · , s̆K gK 2 ,
1 + ρd ηl k 2 [(GH G)−1 ]l,l ηl
gl 2
l=k l=1 and Λ is given in (2).
gk 22 ρu η̆k ρu η̆k Proof: Rewrite the MR uplink SINR from Table I as (3).
Uplink ρu  [(GH G)−1 ]k,k
1+
gk 22 l=k
η̆l |gkH gl |2 Solve (3) for η̆. If η̆ ∈ RK 0+ satisfies η̆∞ ≤ 1, then we have
found a power control that meets the SINR target. Otherwise
the SINR target cannot be met. 
3) Downlink Power Control for ZF Processing:
and Theorem 2.3: For ZF processing, a downlink target SINR
η̆ = (η̆1 , · · · , η̆K )T ∈ RK
0+ vector s = (s1 , · · · , sK )T ∈ RK0+ can be met if and only if
K

is the uplink power control vector with individual power
constraint ρd ≥ sk [(GH G)−1 ]k,k (4)
k=1
η̆∞ ≤ 1.
and the corresponding downlink power control is given by
η = αs
B. Power Controls to Meet Target SINR
K
With scheduling, some users will be served in fewer slots where α ≤ 1/ k=1 sk .
than others. It may be desirable to serve those users with Proof: From Table I, we have
higher rates. In order to be able to assign different target SINR K

to different users, we need to find power controls that can meet ηk = sk ρ−1
d [(GH G)−1 ]l,l ηl = sk α, k = 1, · · · , K,
a given target SINR requirement given by a vector in RK 0+ .
l=1

For two channel vectors gk and gl , we define the channel where α does not depend on k. Since downlink power con-
correlation as straint requires η1 ≤ 1, we have
 
 gk gl  |gkH gl | K
 K

γk,l =  , = .
gk 2 gl 2  gk 2 gl 2 1 ≥ η1 = ηk = α sk .
k=1 k=1
1) Downlink Power Control for MR Processing: K
Thus α ≤ 1/ k=1 sk . In particular, we can choose α =
Theorem 2.1: For MR processing, a downlink target SINR K
1/ k=1 sk . We see that to meet or exceed the target SINR,
vector s = (s1 , · · · , sK )T ∈ RK
0+ can be met if and only if
we must have
[IK − diag(s1 , · · · , sK )Λ] η = ρ−1 ρd η k
d b1 (1) K ≥ sk , k = 1, · · · , K. (5)
−1 ] η
l=1 [(G G)
H
l,l l
where
⎛ ⎞ With ηk = αsk we see that (5) becomes
0 γ1,2
2
··· γ1,K
2
K
⎜ γ2,1 ⎟ 

2
0 ··· ⎟ γ2,K
2
ρd ≥ [(GH G)−1 ]k,k sk
Λ =⎜ . .. .. ⎟ .. (2)
⎝ .. . . ⎠ . k=1

γK,1
2
γK,2
2
··· 0 which is a necessary and sufficient condition that the target

−2 T SINR requirement (s1 , · · · , sK )T can be met. 
b1 = s1 g1 −2 −2
2 , s2 g2 2 , · · · , sK gK 2 4) Uplink Power Control for ZF Processing:
has a solution η ∈ RK Theorem 2.4: For ZF processing, an uplink target SINR
0+ that satisfies η1 ≤ 1. The downlink
power control is given by η. vector s̆ = (s̆1 , · · · , s̆K )T ∈ RK
0+ can be met if and only if

Proof: Rewrite the MR downlink SINR from Table I as (1). ρu ≥ max s̆k [(GH G)−1 ]k,k (6)
k=1,··· ,K
Solve (1) for η. If η ∈ RK 0+ satisfies η1 ≤ 1, then we have
found a power control that meets the SINR target. Otherwise and the corresponding downlink power control is given by
the SINR target cannot be met. 
η̆k = s̆k ρ−1
u [(G G)
H −1
]k,k , k = 1, · · · , K.
2) Uplink Power Control for MR Processing:
Theorem 2.2: For MR processing, an uplink target SINR Proof: From Table I, we have
vector s̆ = (s̆1 , · · · , s̆K )T ∈ RK0+ can be met if and only if
η̆k = s̆k ρ−1
u [(G G)
H −1
]k,k ≤ 1, k = 1, · · · , K.

IK − diag(s̆1 , · · · , s̆K )Λ̆ η̆ = ρ−1
u b̆1 (3) Thus s̆k [(GH G)−1 ]k,k ≤ ρu , k = 1, · · · , K. 

Authorized licensed use limited to: National Tsing Hua Univ.. Downloaded on August 12,2021 at 09:53:07 UTC from IEEE Xplore. Restrictions apply.
2018 IEEE 19th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)

Remark 2.5: From Theorems 2.3 and 2.4, the power controls Algorithm 3.3: Construction of a maximal feasible grouping:
for ZF processing are given explicitly in terms of the SINR Given a K × K hollow symmetric matrix Γ = [γi,j ] with
requirement. For MR processing, however, the power controls γi,j ∈ [0, 1], ∀i, j, and a correlation threshold γh , let γmax =
are given implicitly as the solution to a linear system, as maxi,j {γi,j }.
stated in Theorems 2.1 and 2.2. If there is a “one-parameter” 1) If γmax ≤ γh , then we have a unique maximal feasible
SINR target, i.e., the SINR requirement for the K users can grouping, which has only one element: K1 = K. Done.
be parametrized by a single parameter: sk = sk (δ), s̆k = 2) If γmax > γh , find all the columns in Γ with all its entries
s̆k (δ̆), k = 1, · · · , K, then bisection searches can be car- ≤ γh . Denote the set of the column indices for these
ried out to find the optimal parameter δ  and δ̆  such that columns as K0 . Let K0c = K\K0 = ∅ (See Remark 3.5).
mink {sk (δ)} and mink {s̆k (δ̆)} are maximized. The SINR’s Continue.
for calculating the throughputs in the numerical examples in 3) Randomly pick an i1 ∈ K0c . Construct a user set K(i1 )
Section IV are obtained this way.  that includes i1 as follows:
a) Let {i1 } → K(i1 ) , i.e., add i1 to K(i1 ) .
III. A N A LGORITHM FOR S CHEDULING U SERS
b) Let L(i1 ) = {k ∈ K : γk,j ≤ γh , ∀j ∈ K(i1 ) }. If
Motivated by the results in [13], where it has been shown L(i1 ) = ∅, randomly pick k ∈ L(i1 ) and add it to
that the effective SINR’s are upper bounded by decreasing K(i1 ) . Repeat until L(i1 ) = ∅ (See Remark 3.6).
functions of the maximal channel correlation, we devise a c) Let K(i1 ) ∪ K0 → K(i1 ) , i.e., set K(i1 ) as K(i1 ) ∪ K0 .
scheduler algorithm in the following. 4) Randomly pick i2 ∈ K\K(i1 ) (see Remark 3.7), and con-
The algorithm is to divide the K users into groups such struct a user set K(i2 ) as in Step 3). If K(i1 ) ∪ K(i2 ) = K,
that correlation between users in each group is upper bounded done. Otherwise, randomly pick i3 ∈ K\[K(i1 ) ∪ K(i2 ) ].
by a given threshold and that each group has a maximum Construct a new user set: K(i 3 ) as in Step 3).
number of users (i.e., adding any users into the group will n−1
5) In general,
if K\
∪ j=1 K(ij ) = ∅, randomly pick in ∈
cause the maximum correlation in that group to exceed the
K\ ∪n−1 K
j=1 (ij ) =
 ∅ and construct a new user set K(in )
given threshold) – this is to multiplex as many users as possible
as in Step 3. Repeat until the union of user sets equals
in each time slot.
K.
More precisely, given a K × K hollow (i.e., diagonal
Remark 3.4: K0 is the set of indices of the users whose
elements are all equal to zero) symmetric matrix Γ = [γi,j ],
channel correlations do not exceed γh . Every set Ki in a
let K = {1, · · · , K}. we would like to find J distinct subsets
maximal feasible grouping must include K0 , although it is
Kj ⊂ K such that γk1 ,k2 ≤ γh for any k1 , k2 ∈ Kj , j =
possible that K0 = ∅. In the example given in Remark 3.2,
1, · · · , J and
J we have K0 = ∅. 
K= Kj . Remark 3.5: If γmax > γh , some of the columns of Γ will
j=1 have entries that are > γh , i.e., K0 is a proper subset of K. 
Remark 3.6: The set L(i1 ) depends on K(i1 ) . 
We shall call the quadruple (Γ, K, {K1 , · · · , KJ }, γh ) a feasi-
Remark 3.7: K\K(i1 ) cannot be empty because if γi1 ,i2 >
ble grouping.
γh , then i2 ∈ K\K(i1 ) . 
To achieve maximum multi-user multiplexing gain, we
would like each subset Kj to contain as many elements as A. SINR for Long-Term Equal Throughput
possible. Thus, we make the following definition. Since some users will be served by all time slots while
Definition 3.1: A maximal feasible grouping is a feasible others will have to take turns to be served, the users that are
grouping such that K1 = K or for any k ∈ K\Kj , there exists not served by all time slots will need to be served with a
a k1 ∈ Kj such that γk,k1 > γh . higher rate in order for all users to have the same long-term
Remark 3.2: Obviously, the maximal feasible grouping may data rate. In the following, we shall calculate the required
not be unique. For example, for K = {1, 2, 3, 4}, SINR for those users.
⎛ ⎞ Let J be the number of groups obtained from Algorithm 3.3.
0 0.7 0.1 0.2
⎜ 0.7 0 0.3 0.4 ⎟ If a user belongs to all groups, it is served all the time and
Γ=⎜ ⎝ 0.1 0.3 0 0.6 ⎠
⎟ therefore should be served with the lowest rate. In general, if
0.2 0.4 0.6 0 a user belongs to j groups, meaning that out of every J slots,
the user is served only in j slots, to achieve equal throughput
and γh = 0.5. Let K1 = {1, 3}, K2 = {2, 4} and K1 =
J log2 (1 + SINRJ/J ) = j log2 (1 + SINRj/J )
{1, 4}, K2 = {2, 3}. Then we have many different maxi-
mal feasible groupings. For example, {K1 , K2 }, {K1 , K2 }, we need to have
{K1 , K2 , K2 }, ..., are all maximal feasible groupings. Thus
SINRj/J = (1 + SINRJ/J )J/j − 1
the number of subsets in a maximal feasible grouping is also
not unique.  where SINRJ/J is the SINR for the users who are served all
Below is an algorithm for finding a maximal feasible the time, and SINRj/J is the SINR for the users who is served
grouping. in j slot in every J slots.

Authorized licensed use limited to: National Tsing Hua Univ.. Downloaded on August 12,2021 at 09:53:07 UTC from IEEE Xplore. Restrictions apply.
2018 IEEE 19th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)

If all the users are served with the same SINR, then for a
user who is served only j (≤ J) times in every J slots with
SINRj/J , it is equivalent that the user is served with
SINRJ/J = (1 + SINRj/J )j/J − 1
in every slot.
Remark 3.8: In the above calculations, we have not taken
into account that each user group may have a different
number of users. Under similar channel conditions, groups
with fewer users can usually achieve a higher common SINR.
Under different channel conditions, even groups with the same
number of users will have different common SINR. Due to
channel hardening in Massive MIMO, this variation is evident
only in the slow fading scale. Inclusion of this variation in the
scheduler may possibly be realized with the results in [13]. 
Let
sk = SINRjk /J = (1 + SINRJ/J )J/jk − 1 (7) Fig. 1. MR downlink max-min per user rate comparison for different
schedulers. M = 256, K = 18, ρd = −3 dB, ρu = −20 dB. In the legend:
where jk is the number of slots that the kth user is served in 1st number is the correlation threshold, 2nd number indicates the scheduler
used: 1=Equal throughput; 2=Equal rate per slot. NS = No Scheduling
every J slots. Theorems 2.1 and 2.2 can be used in conjunction
with bisection method to search for the largest downlink
(respectively uplink) SINRJ/J that satisfies the downlink •Base station service antenna gain = 0 dBi
(respectively uplink) power constraints for MR processing. •Base station receiver noise figure = 9 dB
For ZF processing, (4) and (6) can be used to search for the • Mobile antenna gain = 0 dBi
largest achievable downlink and uplink SINRJ/J . • Mobile receiver noise figure = 9 dB
Note that (7) is a “one-parameter” parametrization of sk we compute
with parameter SINRJ/J . See Remark 2.5.
ρd β = 132 dB − 135 dB = −3 dB
IV. N UMERICAL R ESULTS
ρu β = 115 dB − 135 dB = −20 dB
Numerical examples of the scheduler performance are pro-
vided in this section. In the simulations, the service antennas are arranged in a
Assume there are M service antennas at the base station circular array with half-λ (carrier frequency wavelength) arc
and K simultaneous users. The channel vector in LoS for the separation. Using Algorithm 3.3 to group users for service in
kth user is given by each coherence time slot, Figs. 1 and 2 show the CDF (cu-
  2π 2π 2π
T mulative distribution function) of max-min per user effective
gk = βk ei λ r1,k , ei λ r2,k , · · · , ei λ rM,k throughput with MR processing, while Figs. 3 and 4 show the
 2π
 2π 2π
T corresponding CDF’s for ZF processing. The performance of
= βk ei λ r1,k 1, ei λ (r2,k −r1,k ) · · · , ei λ (rM,k −r1,k ) two schedulers with different levels of correlation thresholds
are shown. The two schedulers are
where βk models the large-scale fading, and rm,k is the
Euclidean distance between the mth service antenna and the 1) Scheduler 1: Equal throughput scheduler with target
kth user terminal. SINR described in Section III-A;
Calculation of ρd and ρu : Assuming a suburban morphol- 2) Scheduler 2: Equal rate per slot scheduler with a common
ogy, with 1.9 GHz carrier frequency, 2 km cell radius, 32 m SINR for all users served in the slot.
base station antenna height, and 1.5 m mobile antenna height. In the plots, in the case when Scheduler 2 is used, the user
COST231 propagation model predicts an over-the-air path loss rate is adjusted downward accordingly if a user is not served
of 135 dB. We shall assume the worst case that all the users are by every slot.
at the cell edge, i.e., we have βk ≡ −135 dB, k = 1, · · · , K.
Note that if βk varies from user to user, a lower rate may A. Discussions
be caused by either high channel correlation or small βk . We provide some discussions on the simulation results.
By assuming constant β for all users, lower rate can only • In both downlink and uplink, both MR and ZF processing,
be caused by high channel correlation. we see that scheduling can substantially improve the user
Along with the following assumptions: throughput performance, in both the median throughput
• M = 256, K = 18 and the 90% likely throughput.
• Base station radiated power = 10 W • Scheduler 2 clearly outperforms “No Scheduling” due to
• Mobile radiated power = 200 mW 1) smaller channel correlations among users served in

Authorized licensed use limited to: National Tsing Hua Univ.. Downloaded on August 12,2021 at 09:53:07 UTC from IEEE Xplore. Restrictions apply.
2018 IEEE 19th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)

6FKHGXOHU&RPSDULVRQ058SOLQN each slot; 2) fewer users are served in each slot. Although

16
Scheduler 1 also has these two same advantages over “No
  Scheduling”, it however requires a small number of users

  who are not served in every slot to be served with x times
 higher rate than majority of users. The result is often a
 
 lower median throughput.
 
• For uplink, Scheduler 2 is clearly preferred. Scheduler
3UREDELOLW\



 1 performs poorly due to mobile power limitation. To

 require a certain mobile to deliver x times throughput


of other mobiles, the other mobiles must decrease their
 throughput to satisfy the requirement.
 • For ZF processing, downlink performs quite well even
with up to 95% channel correlation.

• Using 90% likely throughput as the performance indica-

          
tor, Scheduler 2 in conjunction with 0.4 as the channel
3HU8VHU6SHFWUDO(IILFLHQF\ EV+] correlation threshold (blue dashed curves) works well in
all four cases: downlink and uplink, MR and ZF.
Fig. 2. Same as Fig. 1, for MR uplink
V. C ONCLUSIONS
A scheduler is necessary when operating a Massive MIMO
in LoS environment. We have proposed a simple scheduler to
properly assign data rates to users, serve users with similar
channel vectors in separate coherence time slot, while aiming
for fairness and maximal system spectral efficiency. Simula-
tions have shown the effectiveness of the proposed scheduler
for both downlink and uplink, and for both MR and ZF.
R EFERENCES
[1] M. Fuchs, G. D. Galdo, and M. Haardt, “Low-complexity space-time-
frequency scheduling for MIMO systems with SDMA,” IEEE Trans. Veh.
Technol., 56(5):2775-2784, 2007.
[2] V. K. N. Lau, “Proportional fair space-time scheduling for wireless
communications,” IEEE Trans. Commun., 53(8):1353-1360, 2005.
[3] A. Pantelidou and A. Ephremides, “Scheduling in wireless networks”,
Foundations and Trends in Networking, 4(4):421-511, now Publishers
Inc., 2011.
[4] T. Yoo and A. Goldsmith, “On the optimality of multiantenna broadcast
scheduling using zero-forcing beamforming,” IEEE J. Sel. Areas Com-
mun., 24(3):528-541, 2006.
Fig. 3. Same as Fig. 1, for ZF downlink [5] S. E. Hajri, M. Assaad, and G. Caire, “Scheduling in Massive MIMO: user
clustering and pilot assignment,” Proc. 54th Annual Allerton Conference,
September 2016.
[6] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers
of base station antennas,” IEEE Trans. Wireless Commun., 9(11):3590-
3600, 2010.
[7] T. L. Marzetta, E. G. Larsson, H. Yang and H. Q. Ngo, Fundamentals of
Massive MIMO, Cambridge Univ. Press, London 2016.
[8] V. Raghavan, S. Subramanian, J. Cezanne, A. Sampath, O. H. Koymen,
and J. Li, “Single-user versus multi-user precoding for millimeter wave
MIMO systems,” IEEE J. Sel. Areas Commun., 35(6):1387-1401, 2017.
[9] H. Yang and T. L. Marzetta, “A macro cellular wireless network with
uniformly high user throughputs,” Proc. IEEE 80th Vehicular Technology
Conference, Sept. 2014.
[10] B. M. Hochwald, T. L. Marzetta, and V. Tarokh, “Multiple-antenna
channel hardening and its implications for rate feedback and scheduling,”
IEEE Trans. Inf. Theory, 50(9):1893-1909, 2004.
[11] H. Yang and T. L. Marzetta, “Massive MIMO with max-min power
control in line-of-sight propagation environment,” IEEE Trans. Commun.,
65(11):4685-4693, 2017.
[12] H. Q. Ngo, E. G. Larsson and T. L. Marzetta, “Aspects of favorable
propagation in Massive MIMO,” Proc. 22nd European Signal Processing
Conference, Sept. 2014.
[13] H. Yang and T. L. Marzetta, “Max-Min SINR dependence on channel
correlation in line-of-sight Massive MIMO,” Proc. IEEE Globecom,
Fig. 4. Same as Fig. 1, for ZF uplink December 2017.

Authorized licensed use limited to: National Tsing Hua Univ.. Downloaded on August 12,2021 at 09:53:07 UTC from IEEE Xplore. Restrictions apply.

You might also like