On Delay Performance Gains From Network Coding

(Invited Paper)
Atilla Eryilmaz Asuman Ozdaglar and Muriel Médard
Laboratory for Information and Decision Systems Electrical Engineering and Computer Science Department
Massachusetts Institute of Technology Massachusetts Institute of Technology
Cambridge, MA, 02139 Cambridge, MA, 02139
Email: [email protected] Email: {asuman, medard}@mit.edu

Abstract— This paper analyzes the gains in delay performance Side Information (CSI) about the state of receiver channels is
resulting from network coding. We consider a model of file available to the base station and when transmission must be
transmission to multiple receivers from a single base station. carried without such information.
Using this model, we show that gains in delay performance from
network coding with or without channel side information can Our analysis shows that, as well as the already well-
be substantial compared to conventional scheduling methods for understood capacity gains, network coding leads to significant
downlink transmission. improvement in delay performance both with and without CSI.
This is potentially important, since depending on the applica-
I. I NTRODUCTION tion, delay performance may be critical to the satisfaction of
With the introduction of third-generation cellular systems the users. Equivalently, with network coding more users can
over the last decade, there has been both a significant increase be supported with the same delay performance of scheduling.
in the capacity of wireless networks and a growing use of wire- Our paper differs from existing work in this area by ex-
less communication for data transmission. An essential feature plicitly modeling delay performance in file downloads and
of the newly emerging wireless networks is the transmission allowing for transmission without CSI, and to the best of our
of files to multiple (potentially heterogeneous) receivers, as knowledge, presents the first quantification of gains in delay
exemplified by transmission of video or music files. performance resulting from network coding. Previous research
While the most common approach to data transmission has instead focused on either optimal scheduling with time-
builds on the scheduling approach, where information is varying channel conditions (see [13, 14]), or on the capacity
transmitted to one of multiple receivers as a function of gains from network coding (see [11, 5, 6, 12, 10, 15]) under
their channel conditions, it has also been recognized that various different scenarios.
broadcasting to multiple receivers using network coding may The rest of the paper is organized as follows: Section II
be more efficient for utilizing the capacity of the network (c.f. describes the model for downloading (multiple) files to mul-
[1, 8, 7, 9]). Although these throughput gains may appear to tiple receivers and introduces different transmission strategies
imply gains in delay through Little’s law, this is not the case considered. Section III focuses on broadcasting a single file
since coding is performed over large blocks and each packet in and investigates the performance of network coding compared
the block must await the completion of the whole block before to the performance of delay optimal scheduling strategies. We
it can be decoded. Despite considerable practical interest in the show analytically and through simulations that using network
use of network coding in wireless communication systems, coding results in significant gains in delay performance relative
gains in delay performance resulting from network coding to delay optimal scheduling strategies both with or without
relative to traditional scheduling have not been analyzed or CSI. In Section IV, we focus on the multiple unicast scenario.
quantified. The best setting to investigate such gains is the We show that in the presence of CSI, network coding achieves
rateless transmission scenario, where data of fixed length is optimal performance when no coding is allowed across sep-
to be communicated over the channel. In this context, the arate files. Moreover, it provides considerable delay gains
comparison in delay performance between traditional schemes compared to scheduling when CSI is not available. This last
and network coding is performed through the completion times point of significant delay gains is particularly interesting given
of the whole data. that network coding does not provide any capacity gains for the
The objective of this paper is to develop a model to study single hop unicast scenario. We complete with final remarks
delays in file downloads with network coding and quantify the and ideas for future work in Section V.
gains resulting from network coding relative to the traditional
scheduling methods. We consider downlink transmission of II. S YSTEM MODEL
(multiple) files from a single base station to multiple receivers
with varying channel conditions. The varying channel con- In this work, we consider the cellular downlink scenario
ditions are modeled as stochastic changes in ON/OFF state where the base station holds a set of files, F. The set of
of the channel. We analyze the model both when Channel receivers is denoted by N . File f ∈ F is demanded by the
set Nf ⊆ N of receivers1 . We are interested in the minimum where af,k [t] ∈ Fq for each f ∈ F and k ∈ {1, · · · , Kf }. The
average time required to complete the download of all the files transmitter chooses the coefficients {af,k [t]} at every time slot
by all the interested receivers, where the transmissions have t. This mode of transmission will be referred to as the Coding
to be done over a time varying channel. We seek an answer Mode (or simply Coding) henceforth.
to this problem with and without the availability of CSI at Given the above model, we are interested in minimizing the
the base station and with and without the possibility of linear amount of time necessary for all the files to be transmitted to
coding in the manner we will describe. all the interested receivers. We will refer to this metric as the
A given file f ∈ F is composed of Kf packets, where completion time. In particular, we would like to find answers
Packet-k of file f is referred to as Pf,k , which is a vector2 of to the following questions:
length m over a finite field Fq . It is assumed that transmissions • If we restrict ourselves to scheduling policies, then what
occur in time slots, each of which is of duration just long is the policy that minimizes the average completion time?
enough to accommodate a single packet transmission. The • If coding is allowed, then what is the best policy and
channel between the base station and the ith receiver is a what is its delay performance?
randomly varying ON/OFF channel. We let Ci [t] ∈ {0, 1} • How do these two policies’ performances compare?
denote the state of user i’s channel in slot t. We assume that We seek answers to these questions under various scenarios.
Receiver-i successfully receives the packet transmitted at slot In Section III, we focus on the case of a single file demanded
t if Ci [t] = 1, and it cannot receive anything if Ci [t] = 0. by all the receivers (broadcast scenario). Then, in Section IV,
We will take each Ci [t] to be a Bernoulli random variable we will consider the other extreme of each receiver demanding
with mean ci that are independent across time and across a unique file (multiple unicasts scenario). We will discuss
receivers. The channels of different receivers can in general be extensions in Section V.
asymmetric. However, in parts of the subsequent analysis we III. B ROADCASTING A S INGLE F ILE
will restrict our attention to symmetric cases in order to have
In this section, we are concerned with the transmission of
tractable formulations. The presence of CSI implies that the
a single file to all the receivers. Since |F| = 1, we will drop
channel state vector, C[t], is known at the transmitter at the
the subscript f in our notation, and denote Packet-k as Pk
beginning of slot t. The system model is depicted in Figure 1.
and the size of the file as K. We will study the minimum
mean completion time of the file using coding in Section III-
A, where we will observe the asymptotic optimality of coding
P1,K
1
...P 1,2 P1,1
Receiver 1
over all possible strategies. Then, we will characterize the opti-
.. C1 [t]
mal scheduling strategy with and without CSI in Section III-B.
.
Comparison of the findings will be presented in Section III-C
Pf,K ...P f,2 Pf,1 Base C2 [t]
Receiver 2 along with a discussion on the strengths and weaknesses.
f
Station ..
File f (Kf packets) . A. Coding with and without CSI
.. C [t]
. N Receiver N It has been shown in the literature that linear coding is
PF,K
F
. . .P F,2
PF,1 sufficient to achieve the maximum achievable rate for multicast
networks [8]. Noticing that the broadcast scenario is a special
instance of a multicast transmission, we consider the set of
Fig. 1. System model policies where the transmitted packet in slot t is given by
XK
P[t] = ak [t]Pk , with ak [t] ∈ Fq for each k ∈ {1, · · · , K}.
Let us use P[t] denote the packet chosen for transmission
k=1
in slot t. If the base station is not allowed to code, then at We will consider the following randomized strategy (c.f. [4]).
any given slot it must transmit a single packet from one of R ANDOMIZED B ROADCAST C ODING (RBC):
{k=1,··· ,K }
the files. Thus, we have P[t] ∈ {Pf,k }{f ∈F } f . This is While (File is incomplete)
the typical mode of transmission considered in literature. We Pick ak [t] uniformly at random from Fq for each k;
P
will refer to this mode as the Scheduling Mode (or simply Transmit P[t] = K k=1 ak [t]Pk ,;
Scheduling). t ← t + 1;
If coding is allowed, then in a slot, say t, any linear Each receiver collects the incoming packets that they could
combination of the packets can be transmitted. Specifically, receive and then decodes all the packets {Pk }{k=1,··· ,K} when
we have K linearly independent combinations of them are received (c.f.
Kf
XX [4] and references therein). Random linear coding arguments
P[t] = af,k [t]Pk,f , imply that the expected number of slots before K linearly
f ∈F k=1 independent combinations can be collected with RBC is given
by
1 We will occasionally use F, N and N to denote the cardinalities of the
XK
f 1
sets F, N and Nf , respectively. .
2 We will consistently use boldface letters to denote vectors. (1 − (1/q)k )
k=1
This expression can be upper-bounded by Kq/(q−1), which in Pk before it is received by Receiver-i. Then,
turn can be made close to K even with reasonably low values © ª
of q. Thus, for all practical purposes, for a large enough field Y i , max KXki + k
k∈{1,··· ,K}
size q, it is sufficient for each receiver to be active K slots
gives the time slot when Receiver-i receives the whole file.
on average before it can decode the whole file. Notice that
Finally, T2 , max Y i gives the completion time of the
information theoretically it is impossible to send the file with i∈{1,··· ,N }
less than K transmissions, and so RBC asymptotically (in q) algorithm. Its mean is described in the next proposition.
achieves the best possible performance over all strategies.
Proposition 2. Under symmetric channel conditions (i.e. ci =
Another important issue is the overhead related with this
c ∈ (0, 1) for all i), we have
mode of transmission. Coding requires dK log2 qe bits of
∞ h
overhead to contain the coefficients of the associated linear E[T2 ] X ¡ ¢KN i
combination, whereas the packet size is dm log2 qe bits. Thus, =γ+ 1 − 1 − (1 − c)t ,
K t=1
for m >> K, the overhead is negligible. Henceforth, we will
consider this scenario, and ignore the overhead. for some γ ∈ (1/2, 1).
Notice that RBC is not only easy to implement, but also Proof: The upper bound of 1 for γ is due to the fact
requires no knowledge of the channel state vector, and as- that k ≤ K. The lower bound of 1/2 follows from stochastic
ymptotically achieves the smallest mean completion time over coupling arguments and heavily relies on the symmetry of
all policies. We will see in Section III-B that the optimal the channel distributions. In particular, consider a sample path
scheduling policy is much more difficult to characterize, even of the channel state process, ω , (C[1], C[2], · · · ). We use
for the symmetric channel conditions. i(ω) to denote the receiver that was the last to complete the
Next, we find the mean completion time expression for file, and k(ω) to denote the index number of the last packet
RBC. Let us define the random variable Yi as the number that Receiver-i(ω) received. With our earlier notation, Y (ω)
of slots before Receiver-i’s channel is ON K times, for gives the completion time of the file at Receiver-i(ω) under
i = 1, · · · , N. Then,
· we can claim
¸ that the mean completion the given sample path. Also, notice that we have Y (ω) =
i(ω) i(ω)
Xk(ω) (ω)K + k, for some integer Xk(ω) that depends on ω.
time is equal to E max Yi , which is given in the next
i∈{1,·,N } Next, for each sample path ω that leads to k(ω) ∈
proposition. {1, · · · , bK/2c}, we will construct another sample path ω̃
Proposition 1. Let T1 denote the completion time of the opti- that has the same probability of occurrence as ω, but leads
i(ω)
mal coding policy given above. Then, we have to Y (ω̃) = Xk(ω) (ω)K + (K − k(ω)). This implies that
" Ã t µ ¶ !#
X ∞ YN X τ −1 (K + 1)
E[T1 ] = K + 1−
(τ −K) K
c̄i ci , E[Y ] ≥ + KE[max Xki ]. (1)
K −1 2 i,k
t=K i=1 τ =K
0 1 The construction of ω̃ = (C̃[1], C̃[2], · · · ) follows the
where BB@ n CCA gives the number of combinations of size m of n following rule:
m  i(ω)
elements, and c̄i , (1 − ci ).  Cj [rK + (K − l)], if r = Xk(ω) (ω),


Proof: The proof follows from combinatorial arguments
C̃j [rK + l] = j = i(ω),
and is omitted due to space constraints. 
 l ∈ {k(ω), K − k(ω)},

Cj [rK + l], otherwise.
B. Scheduling Mode
In this mode, unlike in the coding mode, the presence or lack It is easy to see that under symmetric conditions this sample
of CSI affects the performance. Hence, these two cases will path has the properties listed above.
be studied separately. Throughout, we will assume symmetric Next, we would like to find the second term in (1). Due to
channels for tractability. i.i.d. assumptions, Xki are also i.i.d. with distribution P(Xki =
1) Scheduling without CSI: To minimize the load of uplink m) = (1 − c)m−1 c, m = 1, 2, · · · . Since this distribution is
transmission which is typically the bottleneck in cellular independent of i and k, we can compute
∞ h
systems, we assume that the transmitter receives feedback X ¡ ¢KN i
from each receiver only at the time when it has just received E[max Xki ] = 1 − 1 − (1 − c)t . (2)
i,k
the whole file. Notice that in this case, all packets have equal t=1

priority. Also, since the channels are independent and identi- The proof is complete once (2) is substituted into (1).
cally distributed (i.i.d.) over time and users, one of the optimal 2) Scheduling with CSI: Before we characterize the optimal
scheduling policies is Round Robin (RR), where Packet-k is scheduling rule with CSI, we demonstrate the suboptimality
transmitted in time slots (mK + k) for m = 0, 1, · · · until all of scheduling compared to coding with the following simple
the receivers get the file. example.
To compute the mean completion time of the above RR Example 1: Consider the case of K = 3 and N = 3,
scheduler, we define Xki to be the number of transmissions of i.e. three packets are to be broadcasted to three receivers.
Consider the channel realizations C[1] = (0, 1, 1), C[2] = M. Hence, we are interested in J ? ([1]N ×K ) where [a]N ×K
(1, 0, 1), C[3] = (1, 1, 0), and C[4] = (1, 1, 1). Thus, in the denotes the all a matrix of dimensions N × K.
first four slots, each receiver can hear the transmission three Before we write the recursion for J ? (M, C), let us define
times. The optimal scheduling rule would transmit P1 , P2 , P3 the function f (·) where M̂ = f (M, C, k) implies that
in the first three slots, leaving Receiver-i in demand for Packet-
i in the fourth slot. Clearly, no scheduling rule can ever com- M̂i,k = Mi,k − Mi,k Ci ∀i ∈ {1, · · · , N },
plete the file download at all three receivers in the fourth slot. M̂i,j = Mi,j ∀i ∈ {1, · · · , N }, j 6= k.
With coding, on the other hand, the following transmissions
will complete the transmissions: (P1 + P2 ), (P2 + P3 ), (P1 + This function describes the next state of the memory matrix
P3 ), (P1 + P2 + P3 ) (see Table I). It is not difficult to see given that Packet-k is served and the channel matrix is C in
that coding will never require more slots than is necessary for the current slot. Then, we can write the following recursion:
scheduling for all other realizations. Hence, we achieve strictly © ª
J ? (M, C) = arg min J ? (f (M, C, k)) + 1{M6=θ} ,
better completion times with coding. k∈{1,··· ,K}

t=1 t=2 t=3 t=4

where 1{A} is the indicator function of the event A.
R1 − P2 |(P2 +P3 ) P3 |(P1 +P3 ) ?|(P1 +P2 +P3 )
The monotone nature of the f (·) function enables us to
R2 P1 |(P1 +P2 ) − P3 |(P1 +P3 ) ?|(P1 +P2 +P3 ) compute J ? (M, C) and π ? (M, C) recursively starting from
R3 P1 |(P1 +P2 ) P2 |(P2 +P3 ) − ?|(P1 +P2 +P3 ) the base state J ? (θ) = 0 (c.f. [2]). This DP formulation
characterizes the optimal policy and its performance, and in
TABLE I. Demonstration of Example 1: Ri corresponds to Receiver-i, ‘−’
denotes OFF channel states, and the entry a|b gives the optimal trans- theory it can be computed starting from a 1 × 1 matrix and in-
missions with scheduling and coding, respectively. With scheduling, no creasing N and K successively. However, as N and K grows,
choice of {Pi } in slot 4 can complete the file at all the receivers for the the necessary number of operations grows exponentially and
given channel realization.
quickly becomes impossible to handle. Thus, we propose an
efficient heuristic policy below and simulate its performance
O PTIMAL S CHEDULING FOR S YMMETRIC C HANNELS : for comparison.
We use Dynamic Programming to find the characterization H EURISTIC P OLICY:
of the optimal scheduling policy for symmetric channel con- We have observed in the above discussions that the optimal
ditions, i.e. ci = c for all i ∈ {1, · · · , N }. Given C[t], the scheduling rule has a complicated structure. Yet, it is possible
scheduler can choose any one of the packets {P1 , · · · , PK } to find practical scheduling algorithms that will perform close
for transmission. A little thought reveals the need of memory to the optimal. Here, we describe a heuristic policy which we
about the previous receptions at each of the receivers. For this believe will provide near optimal performance.
purpose, we define Mi,k [t] to be the memory bit associated At any given time slot t, let us denote the set of nodes with
with Packet-k and Receiver-i. In particular, Mi,k [t] = 1 (or an ON channel (also called the set of active receivers) by
0) implies that Receiver-i has not received (or has received) A[t] , {i ∈ {1, · · · , N } : Ci [t] = 1}. Under the symmetric
Packet-k in the slots 1, · · · , t − 1. Moreover, we will use M[t] conditions that we assumed, the packet that would provide the
k=1,··· ,K
to denote the matrix of memory bits [Mi,k [t]]i=1,··· ,N . most benefit should intuitively be transmitted over the channel.
We let Π denote the set of feasible stationary policies We propose that the benefit of a packet be measured in the
that can be implemented by the base station. Each policy number of nodes in A[t] that has not yet received that packet.
π ∈ Π defines a mapping from the pair (M[t], C[t]) to the The underlying idea is to transfer the maximum number of
set {1, · · · , K} describing the packet to be sent at time t. useful packets over the channel at any given time. These
Note that the policy is stationary in the sense that it is only remarks point to the heuristic algorithm given next.
a function of the matrix and channel conditions at the time.
The i.i.d. nature of the arrivals and departures imply that this H EURISTIC B ROADCAST S CHEDULING (HBS):
is the optimal policy among all policies, including those that If (t = 1)
are time dependent. Mi,k [t] ← 1 for all k ∈!{1, · · · , K}, i ∈ {1, · · · , N };
To characterize the optimal policy we let J π (M, C) = X
K X
N
While Mi,k [t] > 0
E [# slots to reach θ with policy π | M[0] = M, C[0] = C] , k=1 i=1
where θ denotes the zero matrix. Then, J ? (M, C) , K[t] , {k ∈ {1, · · · , K} : ∃i ∈ A[t] with Mi,k [t] = 1};
min J π (M, C) is the minimum completion time of the If (K[t] 6= ∅) X
π∈Π
optimal algorithm if it starts from M and the first channel is T [t] , arg max Mi,k [t];
k∈K[t]
C. Also, π ? (M, C) , arg min J π (M, C) gives the optimal i∈A[t]
π∈Π Pick a k? ∈ T [t];
policy. Mi,k? [t] ← 0 for all i ∈ A[t];
Observe that once we solve J ? (M, C) for all C, we can Transmit Packet-k? over the channel at slot t;
t ← t + 1;
compute J ? (M) , EC [J ? (M, C)] , where the expectation
is over the channel realizations. Thus, J ? (M) denotes the In the algorithm, each packet in K[t] has at least one receiver
mean completion time of the optimal algorithm starting from with an ON channel in slot t which demands that packet.
 Performance of different Schemes for K=30, c=1/2
400 depending of the file sizes and the computation capacity of
the receivers.
350

IV. M ULTIPLE U NICAST T RANSMISSIONS

300

E [T2] Upper bound In this section, we consider the scenario where N receivers
Mean Completion Time

E [T2] Lower bound

250 with symmetric channel conditions demand a unique files, i.e.
E [T ]
1
Heuristic Policy F = N. In this case, it is not clear whether coding will have
200
the dominating behavior as it did in the broadcast scenario.
Again, the availability of CSI is important. We will first study
150
the scheduling case and then move on to the coding case.
100
A. Scheduling for Multiple Unicasts
50
1) Scheduling without CSI: Without CSI, the obvious op-
timal scheduling is again Round Robin, except that it must
0
0 10 20 30 40 50 60 be performed across files and across packets in each file. In
Number of users, N
particular, in the first round the first packet of each file is
Fig. 2. Performance of the two schedulers for a file of length 30 packets transmitted one after another, and in the next round the second
and c = 1/2.
packets are transmitted consecutively. When the end of a file
is reached, we move to the first packet and continue until all
Clearly, those packets that are not in K[t] should not be chosen the packets of a file is received by its receiver. Only then we
for transmission. If K[t] 6= ∅, then we define T [t] to be the remove that file from the RR scheduler and continue with the
set of packets in K[t] that yield the most benefit in slot t. remaining ones.
Then, a packet from T [t] is picked for transmission in slot The mean delay performance of this scheduling rule is easy
t. In our simulations, we considered a random picking of to compute using recursive arguments, which is omitted here
one of the packets in T [t]. However, the performance can be since it does not add any significant insights to our analysis.
slightly improved by using more sophisticated methods. For 2) Scheduling with CSI: Here, the constraint is to serve at
example, for N = 2, the packet picked from T [t] may be most one receiver at every time slot. This problem is a special
chosen amongst those packets that has already been received case of a problem studied by Tassiulas and Ephremides in
by the OFF receiver. Then, every time a receiver is ON, it [13] with no arrivals to the system. The following policy is
will receive a useful packet until all its packets are complete. introduced in [13].
Thus, this algorithm gives the optimal policy for N = 2. L ONGEST C ONNECTED Q UEUE (LCQ):
The generalization of the picking method to general K is t ← 0;
complicated and requires increasing memory to operate. On Qi ← Ki for all i ∈ {1, · · · , N };
the other hand, the complexity of HBS at each iteration of the Do
t ← t + 1;
loop is O(KN ) and requires no extra memory, and hence it i? [t] ← arg max{Ci [t]Qi };
is relatively easy to implement. 1≤i≤N
if(Ci? [t] 6= 0)
C. Comparison Transmit Pi? ,Qi? ;
Qi? ← max(0,! Qi? − 1);
In this section, we perform numerical computations and X
N

simulations to compare the performance of various schemes While Qi > 0 ;

i=1
we have discussed so far. A typical scenario is depicted in Return t; // Completion time
Figure 2, where a file of size 30 packets is to be transmitted
to a varying number of receivers, where each channel is ON In the policy, Qi is used both as a pointer to the index of
or OFF equiprobably at every time slot. the next packet to be transmitted to Receiver-i, and also as
The figure demonstrates the strength of the coding policy the number of packets yet to be transmitted to Receiver-i.
to the scheduling policy with and without CSI. We further Thus, LCQ is a myopic policy that favors the receiver with
observe that as N increases the advantage of using coding the maximum number of packets to be received among all
improves. connected receivers. We repeat the result of [13] for future
reference.
In this section, we have seen that either with or without
CSI, coding provides a considerable gain in the mean delay to Proposition 3 ([13]). Under symmetric channel conditions
download a given file to multiple receivers over a time-varying (i.e. ci = c for all i), LCQ is minimizes the delay over all
medium. Moreover, its operation is significantly easier than the scheduling policies. In other words,
scheduling policy. However, it requires an additional decoding
operation at the receivers, which may or may not be critical T LCQ ¹st T π ,
where T LCQ denotes the completion time under the LCQ Proposition 4. Under the symmetric channel conditions (i.e.
policy and π is any other feasible scheduling policy3 . ci = c for all i ∈ N ), the mean delay minimizing partitioning
This result is very strong and implies that E[T LCQ ] ≤ is obtained when bj = 1 for all j, and the optimal policy is to
E[T π ] for any feasible scheduling policy π. implement LCQ.
B. Coding for Multiple Unicasts Proof: Consider any given partitioning of the files, say
P = {Cj }Jj=1 , and let πP denote the optimal policy for this
A deep understanding of achievable rates for multiple uni- partitioning, which is not known in general. Also, let T πP
cast sessions in a network is still an open problem. In general, be the random variable that denotes the completion time of
it is not clear whether network coding should be performed, all the files under the policy πP . In other words, T πP is the
and if it should what the strategy must be. We will tackle this first slot when each receiver in Class-j received K j linear
problem for the downlink model at hand. combinations of the packets from within their class, for all j.
We define the set of coding classes that partitions F (or
We use ω = (C[1], C[2], · · · ) to denote a sample path of the
equivalently N ) into J subsets. We use Cj to denote the files
channel state process. Notice that the policy and ω determines
(or equivalently receivers) in Class-j. We set the restriction that
T πP (ω).
only those files within the same class will be linearly coded
Next, we will define a new policy π̃ and show that it satisfies
with random coefficients as in RBC, while files of different
T π̃ (ω) ≤ T πP (ω) for all feasible ω. For a given ω, if πP
classes will not be mixed. Notice that for each class, say
serves Class-j in slot t, then π̃ will send only the head-of-
Cj , this P
strategy effectively results in a single file of length
line packet of one of the connected receivers in the same
Kj , f ∈Cj Kf that is demanded by bj , |Cj | distinct class which received the minimum service so far. In other
receivers. Hence, the multiple unicasts problem is converted
words, amongst the connected receivers in Class-j, only the
into a special case of multiple multicasts with each multicast
receiver that has the maximum number of remaining packets
having a disjoint set of receivers. Notice that the description
is served. Notice that this policy does not do any coding, and
of the strategy is yet incomplete, because we must describe
hence requires Receiver-f in Class-j to successfully receive
how to “schedule” the transmissions of different classes. We
Kf packets of its file instead of K j packets as in πP .
will investigate this question with and without CSI.
To see that T π̃ (ω) ≤ T πP (ω), observe that whenever Class-
1) Coding without CSI: In this case, as in Section III-B.1,
j is served under πP , at most one packet (or one degree of
we assume that each receiver informs the base station when
freedom) can be received by each receiver in that class. Thus,
it can decode its own file, which in turn implies that it can
before all of its receivers can decode their own packet, Class-j
decode all the files within its class. The optimal policy is again
must be served at least K j times. But, with π̃ we can send a
going to be of the form of Round Robin over the coding
single degree of freedom to one of the connected receivers in
classes. We will consider the case of bj = b and K j = K̃
Class-j whenever that class is served under πP . Since for each
equal for all j. If J denotes the total number of coding classes,
f ∈ Cj , only Kf degrees of freedom are required for Receiver-
then only a combination from Cj will be transmitted in slot
f with π̃, all the receivers P complete their reception when
(mJ + j) for m = 0, 1, · · · until all the receivers get their
Class-j is served K j = f ∈Cj Kf times. These arguments
files.
prove that for any feasible sample paths the completion of the
Notice that the analysis of the RR scheduler of Section III-
new policy is not larger than that of πP for any partition P.
B.1 does not directly apply to this case, because here once all
To complete the proof, we need to show that T LCQ ¹ T π̃ .
the receivers of a class, say Cj , decode their file, then that class
To that end, we note that π̃ is actually a scheduling policy,
can be extracted from the round robin cycle. Nevertheless,
where at each slot a single packet is transmitted over the
similar analysis based on recursive formulations can be used
channel. Thus, an application of Proposition 3 completes the
for this setting. This analysis is omitted here due to space
proof.
constraints. We remark that without CSI the gain in grouping
subsets of users as described above is only due to the decreas- C. Comparison
ing size of the cycles as groups complete their receptions. If
In this section, we compare the typical performance of
the period of each cycle were kept constant at its starting value
various policies for reasonable parameters. We take bj = b for
of J throughout the operation, then grouping would have no
all j and Kf = K for all f ∈ F. Moreover, we let K = 30 and
effect on the average delay performance, because in such a
N = F = 12 and study the mean completion time behavior
scenario we would be comparing the expected number of slots
of the scheduling and coding strategies with and without CSI.
before K ON channels are observed to 1/b times the expected
Regarding the channel connectivity statistics, we assume that
number of slots before bK ON channels are observed.
ci = 1/2 for all the channels. Figure 3 depicts the simulation
2) Coding with CSI: In the presence of CSI, we must
results of the policies discussed above for varying number of
determine the optimal partitioning of the files {Cj }, and also
classes. In the figure, we observe that the performance of the
find the optimal scheduling policy across these classes. The
LCQ scheduler serves as a lower bound as we have proved in
following proposition finds the optimal policy using stochastic
Proposition 4. Since the optimal coding policy is not described
coupling arguments.
an arbitrary b, in the simulation we use the following heuristic
3¹ is a stochastic ordering as described in [13]. policy: at each time slot among the classes with the maximum
st
 Performance of different Schemes for K=30, N=12
2200 symmetric channel conditions. The assumption of perfect CSI
is unrealistic for actual systems. Instead, if we assume the
2000
availability of CSI for a fraction of the time, then by using
1800
random coding, significant gains can still be achieved, whereas
1600 Scheduling w/o CSI scheduling will be more vulnerable to the lack of CSI.
Sched. with CSI = LCQ
One rule of thumb we obtained from our analysis was to
Mean completion time

Coding w/o CSI

1400
Coding with CSI
code across packets within a file, but to avoid coding across
1200
files. This observation will help us in finding the optimal policy
1000 for the general multiple multicast scenario.
800
An important extension is to consider more general channel
statistics. More interestingly, one can consider a fully con-
600
nected network where each of the N nodes wants to broadcast
400 its own file to the rest of the receivers over time varying
200
channels (see [3] for a variation of this problem).
Another direction is to use our framework to analyze the
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 performance of various strategies under differentiated services.
Number of classes (J = N / b) In particular, we want to specify priority classes with vary-
ing delay constraints, and find the policy that minimizes a
Fig. 3. Mean completion time of various strategies for c = 1/2.
weighted sum of the mean completion times.
R EFERENCES
number of connected receivers, the policy serves the class
with the maximum degrees of freedom yet to be transmitted. [1] R. Ahlswede, Ning Cai, S. R. Li, and R. W. Yeung. Network information
flow. IEEE Transactions on Information Theory, 46:1204–1216, July
This policy, when b = 1 is the same as the LCQ policy. For 2000.
this policy, we observe that the mean delay value achieved [2] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction
decreases to half its value when b is decreased from 12 to to Algorithms. M.I.T. Press, McGraw-Hill Book Company, London,
England, 2001.
1. We also observe that in agreement with our arguments, [3] S. Deb and M. Médard. Algebraic gossip: A network coding approach
the performance of the coding without CSI improves as b to optimal multiple rumor mongering. In Proceedings of the 43rd An-
decreases, but this decrease is rather insignificant. nual Allerton Conference on Communication, Control and Computing,
October 2004.
Without CSI, the performance of scheduling is significantly [4] T. Ho. Networking from a network coding perspective. PhD thesis,
worse than the coding solution. In this particular case, we MIT, 2004.
observe almost a threefold delay with scheduling as opposed [5] D. Katabi, S. Katti, W. Hu, H. Rahul, and M. Médard. The impor-
tance of being opportunistic: On practical network coding for wireless
to coding. Given that the single-hop multiple unicasts scenario environments. In Proceedings of Annual Allerton Conference on Com-
does not improve the capacity of the channel, the presence of munication, Control and Computing, October 2005.
such a considerable delay gain is particularly striking. [6] D. Katabi, S. Katti, W. Hu, H. Rahul, and M. Médard. On practical
network coding for wireless environments. In Proceedings of IEEE
The fact that both with and without CSI the performance of International Zurich Seminar on Communications, 2006.
the coding strategy improves as b goes to one implies that for [7] R. Koetter and M. Médard. Beyond routing: An algebraic approach to
unicast transmissions, it is best to code within files, but not network coding. IEEE Transactions on Information Theory, 11:782–795,
October 2003.
across them. [8] S.-Y. R. Li, R. W. Yeung, and Ning Cai. Linear network coding. IEEE
Transactions on Information Theory, 49:371–381, February 2003.
V. D ISCUSSIONS AND F UTURE D IRECTIONS [9] D. S. Lun, M. Médard, R. Koetter, and M. Effros. Further results on
coding for reliable communication over packet networks. In Proceedings
In this work, we have introduced a simple model where of IEEE International Symposium for Reliable Communication over
delay performance of network coding can be investigated Packet Networks, pages 1848–1852, 2005. (An extended version has
and compared to the traditional method of scheduling. Under been submitted to the IEEE Transactions on Information Theory).
[10] D. S. Lun, P. Pakzad, C. Fragouli, M. Médard, and R. Koetter. An
various scenarios, we have identified the optimal policies and analysis of finite-memory random linear coding on packet streams. In
derived analytical expressions for the delay expressions. We Proceeding of NetCod, 2006.
observed that with easily implementable coding strategies, [11] D. S. Lun, N. Ratnakar, R. Koetter, M. Médard, E. Ahmed, and H. Lee.
Achieving minimum cost multicast: A decentralized approach based on
significant delay gains can be obtained. network coding. In Proceedings of IEEE Infocom, pages 1607–1617,
We have analyzed two scenarios: the case when all receivers 2005.
demand a single file (broadcast case), and the case when [12] S. Shakkottai P. Gupta S. Bhadra. Min-cost selfish multicast with
network coding. In Proceeding of NetCod, 2006.
each receiver demands a different file (multiple unicasts case). [13] L. Tassiulas and A. Ephremides. Dynamic server allocation to parallel
Under both scenarios, when the channel side information is not queues with randomly varying connectivity. IEEE Transactions on
available we have observed the advantage of using random Information Theory, 39:466–478, March 1993.
[14] B. S. Tsybakov. File transmission over wireless fast fading downlink.
coding strategies over pure scheduling approaches. However, IEEE Transactions on Information Theory, 48:2323–2337, August 2002.
when CSI is present, it turned out that coding gives consid- [15] Yunnan Wu. Network coding for multicasting. PhD thesis, Princeton
erable gains for the broadcast scenario, whereas scheduling University, November 2005.
is the best policy for the multiple unicasts scenario under