588 IEEE COMMUNICATIONS LETTERS, VOL. 23, NO.

4, APRIL 2019

An Intelligent QoS Algorithm for Home Networks

Wen-Jyi Hwang , Tsung-Ming Tai, Bo-Ting Pan, Tun-Yao Lou, and Yun-Jie Jhang

Abstract— A novel quality of service (QoS) management

algorithm for home networks is presented in this letter.
The algorithm is based on service prediction for intelligent
QoS management. The service prediction is carried out by a
general regression neural network with a profile containing the
past records of the service. A novel profile updating technique
is proposed so that the profile size can remain small for fast Fig. 1. An example of heterogeneous home network.
bandwidth allocation. The analytical study and experiments over
real LAN reveal that the proposed algorithm provides reliable
QoS management for home networks with low computational The General Regression Neural Network (GRNN) [7] has
overhead. been found to be promising for service response prediction [8]
without offline training. However, a drawback of the GRNN
Index Terms— Quality of service, neural networks, home
networks. algorithm in [8] is that the computation complexities may be
high for a frequently requested service. This is because the
I. I NTRODUCTION
service response prediction is based on a profile consisting

T HE coexistence of multiple wireless and wireline

network resources is an inherent characteristic in a
heterogeneous home network [1]. Furthermore, with the pro-
of past records of the service. The size of the profile grows
with the service usage. This may introduce large computation
overhead for QoS management.
liferation of network applications and appliances, diversi- The objective of this study is to present a novel
fied quality are usually desired for communication services. GRNN-based QoS management algorithm for home networks.
A Quality of Service (QoS) algorithm for the services well- The GRNN is used for the service response prediction based
suited for the heterogeneous nature of home networks will then on a profile of past delivery records. A novel profile updating
be beneficial for efficient utilization of network resources. algorithm is proposed so that the profile size of a service
A solution to the QoS problem in a home network may could remain small as the service usage grows. In this way,
be based on approaches such as integrated services or multi- computation complexities may still be low for frequently
protocol label switching protocols [2]. However, bandwidth used services. Analytical studies and experimental results
subscription may be required by these methods. It could confirm the superiority of the proposed algorithm over existing
result in large computational overhead for enforcing the sub- approaches for QoS management of home networks.
scription. A number of optimal tracking algorithms [3], [4]
II. T HE P ROPOSED A LGORITHM
for networked control systems are also beneficial for the
QoS management. Nevertheless, the systems are assumed to Figure 1 shows a home network example containing a
be time-invariant. Effective QoS management may then be number of domains. Each domain consists of a bridge, which
difficult when the characteristics of networks are time-varying. forwards data from an appliance in that domain to the other
An alternative is to adopt service response prediction for domains. The proposed algorithm is deployed in the bridges
QoS management. Given a bandwidth allocation, a service for the QoS management. Home networks are considered in
response prediction scheme forecasts the corresponding ser- this study only as examples. The QoS management for other
vice quality. The QoS management is then carried out from LAN networks such as public/office networks can also be
the prediction results subject to a desired transmission quality. implemented by the proposed algorithm for more applications.
Although some existing service prediction techniques [5], [6] Consider a typical delivery scenario in a home network,
are effective, offline training is usually required. It may then where a service delivering data from an apparatus in domain
be difficult to adopt the training results for home network s to an apparatus in domain t is carried out. The corresponding
environment with time-varying nature. bridges are connected by n links. Let xj , j = 1, . . . , n,
be the bandwidth of link j  allocated for the service, and
n
Manuscript received December 9, 2018; revised January 30, 2019; accepted x = {x1 , . . . , xn }. Let |x| = j=1 xj be the total bandwidth
January 31, 2019. Date of publication February 4, 2019; date of current
version April 9, 2019. This work is supported by the Ministry of Science
of the bandwidth allocation x. Let R be the source data rate of
and Technology, Taiwan, under Grant MOST 107-2221-E-003-001-MY2. The the service. When |x| ≥ R, let Residual Allocation Bandwidth
associate editor coordinating the review of this letter and approving it for (RAB) be RAB = |x|−R. Moreover, when |x| < R, we define
publication was A. Mellouk. (Corresponding author: Wen-Jyi Hwang.)
W.-J. Hwang, B.-T. Pan, T.-Y. Lou, and Y.-J. Jhang are with the Depart-
Data Loss Rate (DLR) as DLR = R−|x|. Both RAB and DLR
ment of Computer Science and Information Engineering, National Tai- can be measured by bridges. A single measurement, termed
wan Normal University, Taipei 117, Taiwan (e-mail: [email protected]; Extended RAB (ERAB), can then be derived as
[email protected]; [email protected]; [email protected]). 
T.-M. Tai is with NVIDIA Corporation, Taipei 114, Taiwan (e-mail: RAB when |x| ≥ R
[email protected]). ERAB = (1)
Digital Object Identifier 10.1109/LCOMM.2019.2897567 -DLR when |x| < R.
1558-2558 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Nvidia Corp. Downloaded on July 22,2020 at 08:34:41 UTC from IEEE Xplore. Restrictions apply.
HWANG et al.: INTELLIGENT QoS ALGORITHM FOR HOME NETWORKS 589

Consider a set of thresholds {η1 , . . . , ηL−1 } satisfying ηi < ηj Algorithm 1 Proposed QoS Management Algorithm
for i < j. Let interval Ik = (ηk−1 , ηk ], for k = 2, . . . , L − 1, Require: initial T = {xi , yi , i = 1, . . . , p}, S, q, and aq .
and I1 = (−∞, η1 ], IL = (ηL−1 , ∞). Let a service response 1: Initialize the optimal bandwidth allocation xOq by (4).
level y be a quantized ERAB, given by
2: loop
y = k when ERAB ∈ Ik , (2) 3: if service required then
4: Carry out the service based on xOq .
where 1 ≤ k ≤ L, and L is the number of response levels. A
QoS level can be specified for a service in terms of service 5: Measure the ERAB defined in (1).
responses. Let Q be the total number of QoS levels available 6: Compute y from ERAB by (2).
for a service. An integer aq is assigned to each QoS level 7: {xp+1 , yp+1 } ← {xOq , y}.
q, q = 1, . . . , Q. When a service is delivered with QoS 8: P ROFILE _U PDATE(T , xp+1 , yp+1 , S, p, aq )
level q, its service response level is expected to be higher than 9: Compute new xOq based on the new T by (4).
(or equal to) aq , where ai < aj for i < j. We can then observe 10: end if
from (2) that a service at the service level q is expected to have 11: end loop
ERAB larger than η(aq −1) . The expected ERAB can be viewed
as the additional reserved bandwidth for a service. Higher QoS
levels have larger expected ERABs so that packet losses are Algorithm 2 Profile Updating Algorithm
less likely when the source data rate surges unexpectedly. 1: procedure P ROFILE _U PDATE (T , xp+1 , yp+1 , S, p, aq )
Our goal is to find an optimal bandwidth allocation x based 2: if p < S then
on a profile T = {xi , yi , i = 1, . . . , p} by the GRNN, where 3: T ← T ∪ {xp+1 , yp+1 }, p ← p + 1.
xi is a bandwidth allocation, and yi is its service response. 4: else
Given x and T , the GRNN first computes
p 5: Find i∗ by (5), xi∗ ← xp+1 , yi∗ ← yp+1 .
yi W (x, xi ) end if
y∗ = i=1 p , (3) 6:
i=1 W (x, xi ) 7: return T , p.
8: end procedure
where W (x, xi ) = exp( −D(x,x σ2
i)
), and D(x, xi ) is the
squared distance between x and xi . Let ŷ be the prediction of
the service response, which is obtained by rounding y∗ in (3).
In the cases the rounded number is less than 1 or larger than lower or maintain the total bandwidth for the resources
L, we set ŷ = 1 and L, respectively. consumption reduction. Consider a service with QoS level q.
Let Bj is the maximum bandwidth at the link j. The search We first note that the set Oq can be rewritten as
space for bandwidth allocation is given by B = {x : xj = Oq = {x : y∗ ≥ aq − 1/2}. (6)
cΔ, 0 ≤ xj ≤ Bj }, where Δ > 0 is the search step size, c ≥ 0
is an integer. Let Oq be the set of the bandwidth allocations From (3) and (6), we see that y∗ and Oq are dependent on
x predicted to be at QoS level q. It can be found by Oq = the profile size p. Let Oq (p) be the set Oq with profile size p.
{x : ŷ ≥ aq }. Let xOq be the bandwidth allocation satisfying Substituting (3) to (6), we obtain
p
xOq = min |x|. (4) yi W (x, xi ) 1
x∈Oq Oq (p) = {x : i=1 p ≥ aq − }. (7)
i=1 W (x, xi ) 2
The allocation xOq is then used for the service in the proposed
It can be shown that (7) is equivalent to
algorithm. Algorithm 1 summarizes the operations of the pro-
posed algorithm. Algorithm 2 is used for the profile updating. Oq (p) = {x : C1 + C2 ≥ C3 }, (8)
Let S be the maximum profile size. When the current profile L 
size p is less than S, the new bandwidth allocation xp+1 and where C1 = u=aq (u − aq )W (x, xi ), C2 =
1
p ai∈I u
q −1

new response yp+1 are appended to the profile, and the size 2 i=1 W (x, xi ), C3 = u=1 i∈Iu (aq −u)W (x, xi ), and
p is incremented by 1. When p reaches S, the profile T is Iu = {i : 1 ≤ i ≤ p, yi = u}. Note that all the terms C1 , C2
updated by replacing {xi∗ , yi∗ } with {xp+1 , yp+1 }, where and C3 in (8) are positive.
 We first investigate the impact on bandwidth allocation
argmini∈P D(xp+1 , xi ) if yp+1 ∈ N , after appending a new record {xp+1 , yp+1 } to the profile T .
i∗ = (5)
argmini∈N D(xp+1 , xi ) otherwise, Two scenarios are considered separately: yp+1 ≥ aq and
yp+1 < aq . In the scenario with yp+1 ≥ aq , a record with
and N = {1, . . . , aq − 1} and P = {aq , . . . , L} are the set of positive response is received. It can be shown that
negative responses and positive responses, respectively.
1
Oq (p+1) = {x : C1 +C2 +(yp+1 −aq + )W (x, xp+1 ) ≥ C3 }.
2
III. Q O S-AWARE BANDWIDTH A LLOCATION (9)
A major feature of the proposed QoS algorithm is that it
Clearly, with yp+1 ≥ aq , all the terms in (9) are positive.
is QoS aware. The algorithm will increase the total band-
Consequently, by comparing (8) with (9), it follows that
width to a service for delivery quality improvement for
a negative response. Alternatively, a positive response will Oq (p + 1) ⊇ Oq (p), when yp+1 ≥ aq . (10)

Authorized licensed use limited to: Nvidia Corp. Downloaded on July 22,2020 at 08:34:41 UTC from IEEE Xplore. Restrictions apply.
590 IEEE COMMUNICATIONS LETTERS, VOL. 23, NO. 4, APRIL 2019

p+1
From (4) and (10), we get Because 0 < W (x, xi ) ≤ 1, it follows that i=1 W (x, xi ) →
∞ as p → ∞. Therefore, Δy∗ → 0 as p → ∞. That is,
|xOq (p+1) | ≤ |xOq (p) |, when yp+1 ≥ aq . (11)
we can decrease the variation of y∗ by increasing profile size
Therefore, after appending a new record yp+1 ≥ aq , our p. This will result in lower variation in bandwidth allocation
algorithm lowers the total bandwidth for the next service. For xOq. To enforce steady bandwidth allocation, a lower bound
the scenario yp+1 < aq , it can be shown that on pi=1 W (x, xi ) can be pre-specified. This is equivalent to
imposing a lower limit on profile size.
1
Oq (p+1) = {x : C1 +C2 ≥ (aq −yp+1 − )W (x, xp+1 )+C3 }.
2
(12) IV. E XPERIMENTAL R ESULTS
The proposed algorithm has been deployed in a real
Because yp+1 < aq , all terms in (12) are positive. We then
LAN network for performance evaluation. As shown in
conclude from (8) and (12) that Oq (p+1) ⊆ Oq (p). Therefore,
Figure 2, there are three bridges in the LAN, namely bridges
|xOq (p+1) | ≥ |xOq (p) |, when yp+1 < aq . (13) 1, 2 and 3. Two Gigabit Ethernet links are connected between
bridges 1 and 2. A link aggregation technique is employed for
We next consider the cases where an old record is removed the combination of the links. Given a service with source data
from the profile. Let {xi∗ , yi∗ } be the record to be deleted. rate R, a data distribution scheme well suited to the proposed
When a record with positive response is removed, yi∗ ≥ aq . algorithm is then employed. Let rj be the source data rate
It can then be shown that assigned to the jth link. The data distribution is then based
1 on rj = Rxj /|x|. The Wi-Fi is used for the connection of
Oq (p−1) = {x : C1 +C2 ≥ (yi∗ −aq + )W (x, xi∗ )+C3 }.
2 bridge 3 to the other two bridges. Each bridge is implemented
(14) by Raspberry Pi 3 B+. In each bridge, the Open vSwitch [9]
Because all terms in (14) are positive, it can be derived from is deployed for packet shaping and traffic queuing. The source
(8) and (14) that Oq (p − 1) ⊆ Oq (p), and data packets of the target service are produced by iPerf [10].
They are delivered by user datagram protocol (UDP). The
|xOq (p−1) | ≥ |xOq (p) |, when yi∗ ≥ aq . (15) maximum throughput provided by the Ethernet adapter of each
When a negative response is removed (yi∗ < aq ), we see that bridge is 300 Mbps. Up to seven 40-Mbps services can be
delivered by the proposed system.
1 The proposed algorithm is deployed in bridge 3 of the
Oq (p−1) = {x : C1 +C2 +(aq −yi∗ − )W (x, xi∗ ) ≥ C3 }.
2 LAN. It first receives ERAB reports from the other bridges.
(16) From the collected information, it then computes the new
Again, all the terms are positive in (16). Therefore, we bandwidth allocation, which is sent to bridge 1 for packet
conclude from (8)(16) that Oq (p − 1) ⊇ Oq (p), and shaping operations by Open vSwitch. The OpenFlow protocol
is used for the bandwidth information delivery. The packet
|xOq (p−1) | ≤ |xOq (p) |, when yi∗ < aq . (17) shaping operations are based on token bucket algorithm, where
As shown in Algorithm 2, when profile size p < S, its a dedicated token bucket is assigned to each link. The token
profile updating process is simply a record appending opera- rate of a token bucket associated with a link is identical to the
tion. Otherwise, the updating process is a record replacement allocated bandwidth to that link.
operation, which can be viewed as an appending operation The evaluation of the two-link system shown in Figure 2
followed by a removal one. Therefore, from (11)(13)(15) for QoS management for 40 transmissions is revealed in
and (17), it can be shown that Algorithm 1 is QoS-aware. Figure 3. In the experiment, the network system has three QoS
The selection of maximum and minimum profile size can levels (i.e., Q = 3) and twelve response levels (i.e., L = 12).
be dependent on the computation complexities for bandwidth We set {a1 , a2 , a3 } = {7, 9, 11} and {η1 , . . . , η11 } =
allocation, and the variation of prediction y∗ in (3), respec- {−11.25, −8.75, −6.25, −3.75, −1.25, 1.25, 3.75, 6.25, 8.75,
tively. It can then be derived from (3) that the computation time 11.25, 13.75} (in Mbps). The ERAB at QoS levels 1, 2 and
then grows linearly with the profile size p. Consequently, when 3 are then expected to be larger than 1.25 Mbps, 6.25 Mbps
the limitation on computation time is imposed, the maximum and 11.25 Mbps, respectively. The initial profile T contains
profile size S can then be determined. One way to find the 16 records. The maximum profile size is S = 31 records. That
limitation is from the usage rate of the bridge responsible for is, for the first 15 transmissions, only the record appending
the computation of bandwidth allocation. When the usage rate operations are carried out for updating the profile. The record
is above a pre-specified threshold, the available computation replacement is then executed afterwards. The maximum
capacity for further extension of the profile may be small. bandwidth at link 1 and link 2 for bandwidth allocation for
In this case, the computation time limitation is reached. a single service are B1 = 50 Mbps and B2 = 30 Mbps,
Let y∗ (p) be y∗ in (3) when the profile size is p. Let Δy∗ = respectively.
|y (p + 1) − y∗ (p)| be the variation of y∗ when the profile
∗ The source data rate R and the total allocated bandwidth
size grows from p to p + 1. It can be shown from (3) that for each QoS level are shown in Figure 3. An additional
background service is also transmitted with source data rate
L−1
Δy∗ ≤ p+1 . (18) 40 Mbps. In this way, the effectiveness of the QoS manage-
i=1 W (x, xi ) ment in the presence of other data packets can be observed.

Authorized licensed use limited to: Nvidia Corp. Downloaded on July 22,2020 at 08:34:41 UTC from IEEE Xplore. Restrictions apply.
HWANG et al.: INTELLIGENT QoS ALGORITHM FOR HOME NETWORKS 591

Fig. 2. A real LAN network for performance evaluation.

Fig. 4. Comparisons of bandwidth allocation results of various

implementations at QoS level 2.

TABLE I
T HE C OMPARISONS OF VARIOUS I MPLEMENTATIONS ON RAB, DLR,
C OMPUTATION T IME AND BANDWIDTH VARIATION AT Q O S L EVEL 2

Fig. 3. The source data rate and total bandwidth allocation of the proposed
algorithm at three QoS levels.

We can see from Figure 3 that the proposed algorithm provides

fast tracking to R at all QoS levels. This is because the
proposed algorithm is QoS-aware. It is able to automatically
increment or decrement the allocated bandwidth to track R
based on the measured ERAB. Furthermore, the measured proposed algorithm. Analytical studies reveal that the proposed
average ERABs associated with QoS levels 1, 2 and 3 for algorithm is QoS-aware. Moreover, experiments over real LAN
the first 40 transmissions are 0.79 Mbps, 6.69 Mbps and networks exhibit that the proposed algorithm offers fast adap-
13.06 Mbps, respectively. The expected ERABs are close to tation with low computational complexities to the variations of
the measured ones. the source data rates. All the results reveal that the proposed
To further evaluate the proposed algorithm, Figure 4 reveals algorithm is able to provide effective QoS management for
the total bandwidth allocation of various implementations at home networks.
QoS level 2. Table I then shows the average RAB, average
DLR, the computation time of bandwidth allocation at R EFERENCES
the final transmission, and the average variation of total [1] C. Cano, A. Pittolo, D. Malone, L. Lampe, A. M. Tonello, and
bandwidth for two successive transmissions. It can be A. G. Dabak, “State of the art in power line communications: From
observed from Figure 4 that the proposed algorithm and the applications to the medium,” IEEE J. Sel. Areas Commun., vol. 34,
no. 7, pp. 1935–1952, Jul. 2016.
algorithm in [8] provide similar total bandwidth allocations [2] M. Marchese, QoS Over Heterogeneous Networks. New York, NY, USA:
for all the transmissions. However, Table I shows that the Wiley, 2007.
algorithm in [8] has higher computation time due to larger [3] X.-S. Zhan, Z. H. Guan, X.-H. Zhang, and F.-S. Yuan, “Optimal tracking
performance and design of networked control systems with packet
profile size. Moreover, the algorithm in [5] based on k Nearest dropouts,” J. Franklin Inst., vol. 350, no. 10, pp. 3205–3216, 2013.
Neighbors (kNN) prediction may have slower source rate [4] X.-S. Zhan, J. Wu, T. Jiang, and X.-W. Jiang, “Optimal performance
tracking capability. We can also see from Table I that the of networked control systems under the packet dropouts and channel
noise,” ISA Trans., vol. 58, pp. 214–221, Sep. 2015.
average DLR and bandwidth variation of the proposed [5] M. Venkataraman, M. Chatterjee, and S. Chattopadhyay, “Evaluating
algorithm become lower as profile size S increases at the quality of experience for streaming video in real time,” in Proc. IEEE
expense of higher computation time. Because variation in Global Telecommun. Conf., Honolulu, HI, USA, Mar. 2009, pp. 1–6.
[6] J. Deng, L. Zhang, J. Hu, and D. He, “Adaptation of ANN based video
response prediction lowers as profile S increases, as shown stream QoE prediction model,” in Advances in Multimedia Information
in (18), the variation in bandwidth allocation lowers as well. Processing (Lecture Notes in Computer Science), vol. 8879, Zürich,
The proposed algorithm then provides flexibilities on both Switzerland: Springer-Verlag, 2014, pp. 313–322.
[7] D. F. Specht, “A general regression neural network,” IEEE Trans. Neural
the computational complexities and variation in bandwidth. Netw., vol. 2, no. 6, pp. 568–576, Nov. 1991.
Smaller profile sizes result in faster computation, while larger [8] W.-J. Hwang, T.-M. Tai, Y.-J. Jhang, Y.-C. Tung, C.-H. Ho, and
ones have more steady bandwidth allocation. S.-Y. Kuo, “Quality of service management for home networks using
online service response prediction,” IEEE Internet Things J., vol. 4,
no. 5, pp. 1773–1786, Oct. 2017.
V. C ONCLUSION [9] OVS: Open vSwitch. Accessed: Jun. 1, 2018. [Online]. Available:
http://openvswitch.org/
Analytical studies and experiments over real home net- [10] iPerf. The Ultimate Speed Test Tool for TCP, UDP and SCTP. Accessed:
works are provided to illustrate the effectiveness of the Jun. 1, 2018. [Online]. Available: https://iperf.fr/

Authorized licensed use limited to: Nvidia Corp. Downloaded on July 22,2020 at 08:34:41 UTC from IEEE Xplore. Restrictions apply.