Int. J. Communications, Network and System Sciences, 2011, 4, 351-355
doi:10.4236/ijcns.2011.45040 Published Online May 2011 (http://www.SciRP.org/journal/ijcns)
New Approach of QoS Metric Modeling on Network
on Chip
Salem Nasri1,2
1CES Laboratory, ENIS, Tunisia
2College of Computer, Qassim University, Qassim, Kingdom of Saudi Arabia
Received February 28, 2011; revised March 21, 2011; accepted April 12, 2011
Abstract
This paper presents a new NoC QoS metrics modeling shaped on mesh architecture. The new QoS model is
based on the QoS parameters. The goal of this work is to quantify buffering requirements and packet switch-
ing techniques in the NoC nodes by analyzing some QoS metrics such as End-to-End delays (EEDs) and
packet loss. This study is based on simulation approach of a 4 × 4 mesh NoC behavior under multimedia
communication process. It proposes a study of NoC switching buffer size avoiding packet drop and mini-
mizing EED. Mainly, we focus on percent flit losses due to buffer congestion for a network loading. This
leads to identify the optimal buffer size for the switch design. The routing approach is based on the Worm-
hole Routing method.
Keywords: NoC, QoS, End-to-End Delay, Packet Loss, Throughput
1. Introduction
According to ITRS, in 2018, ICs will be able to integrate
billions of transistors, with feature sizes around 18 nm
and clock frequencies near to 10 GHz [1]. In this context,
a network on chip (NoC) appears as an attractive solution
to implement future high performance networks and
more suitable QoS managements. A NoC is composed by
IP cores and switches connected among them by com-
munication channels [2]. End-to-end communication
system is accomplished by the exchange of data among
IP cores. Often, the structure of particular messages is
not adequate for the communication purposes. This leads
to the concept of packet switching. In the context of
NoCs, Packets are composed by header, payload, and
trailer. Packets are divided into small pieces called Flits
[3,4]. It appears of importance, to meet the required per-
formance in NoC hardware resources. It should be speci-
fied in an earlier step of the system design the main at-
tention should be given to the choice of the physical
buffer size in the node. The EED and packet loss are
some of the critical QoS metrics. Some real-time and
multimedia applications bound up these parameters and
require specific hardware resources and particular man-
agement approaches in the NoC switch. The best case is
to provide the shortest constant EED or at least with the
minimum fluctuation [5,6].
The paper is organized as follows. Section 2 presents
the network on chip internal architecture and network
routing packets. Section 3 introduces the notion of QoS
metric modeling based on the QoS parameters. Simula-
tion results for the NoC architecture target are presented
and discussed to bring out some physical requirements
enabling QoS metric evaluation based on the QoS pa-
rameters for one class of application in the Section 4. We
finish by the conclusions and perspectives.
2. NoC Architecture and Packet Routing
NoC topologies are defined by the switches connection
structure. The studied NoC architecture assumes each
switch has a set of bi-directional ports linked to its
neighbor switches and to an IP core. It is built on 4 × 4
mesh topology as shown in Figure 1.
Each switch has routing control unit and five
bi-directional ports: East, West, North, South, and Local.
Each port has an input buffer for temporary information
storage. The local port establishes a communication be-
tween the switch and its IP core. The other ports are con-
nected to the neighbor switches. The routing control unit
implements the logic arbitration and packet-switching
algorithm. The main critical parameters driving the switch

S. NASRI 352
Figure 1. 4 × 4 mesh NoC structure.
performances are the memory access time (reading and
writing) and the transition time through the switch (from
input to output). In fact, it is important to minimize data
bufferisation time because it reduces flits throughput,
increases EED, causes jitter, and can lead to data loss if
there is insufficient memory space to store all incoming
data flows waiting to be transmitted. Theses communica-
tion parameters must be considered together with hard-
ware system constraints related to circuit area and com-
puting frequency optimization [7].
The main part of the switch is the flit scheduler. It is
based on the Deficit Weighted Round Robin (DWRR)
technique for the management of the data queuing. In
this technique the switch defines many application
classes and it associates a weight to each class. The
switch bandwidth is then divided to input traffic classes
according to their bandwidth requirement.
The scheduling approach managing the output switch
buffer based on DWRR defines mainly two parameters:
 The Counter which specifies the total number of
bytes that the queue is permitted to transmit at each
time it is visited by the scheduler.
 A quantum of service proportional to the weight of
the queue, it is expressed in bytes.
The Counter for a queue is incremented by the quan-
tum value each time the queue is visited by the scheduler.
In the DWRR the scheduler algorithm starts by deter-
mining the number of bytes at the head of the queue.
Counter = counter + quantum
Data in the queue is sent only if the size of the packet at
the head of the queue is less than or equal to the variable
Counter. The variable Counter is reduced by the number
of bytes being sent and data is transmitted on the output
port. The scheduler continues to send data from this queue
until data in the queue is less than the value of Counter or
the queue is empty. In this case the variable Counter will
be set to zero. Then the scheduler moves on, to serve the
next non-empty queue [8].
3. QoS Requirements in NoC Design
3.1. QoS Tentative Definition
The Quality of Service (QoS) refers to a broad collection
of networking technologies and parameters. The goal of
QoS is to provide guarantees on the ability of a network
to deliver predictable performances. Elements of network
performance within the scope of QoS often include
availability (uptime), bandwidth (throughput), latency
(delay), and error rate. QoS involves, also, prioritization
of network traffic classes. It can be targeted at a network
interface, toward a given server or router's performance.
In terms of specific applications a network monitoring
system must typically be deployed as part of QoS, to
insure that networks are performing at the desired service
level [9]. In packet-switched networks it refers to the
probability of the network meeting a given traffic con-
tract [10].
3.2. QoS Parameters
QoS is especially important for the new generation of
Internet applications such as VoIP, video-on-demand and
other consumer services. Some core networking tech-
nologies like Ethernet were not designed to support pri-
oritized traffic or guaranteed performance levels, making
much more difficult the QoS implementation solutions.
In communication networks, such as Ethernet, through-
put is the average rate of successful packets delivery over
a communication channel. People are often concerned
about measuring the maximum data throughput rate of a
communications link or network access. A typical simple
method of performing a measurement is to transfer a file
F and measure the time T taken to do so.
EED concerns the time for a packet to reach its desti-
nation, because it gets held up in long queues, or takes a
more indirect route to avoid congestion. Alternatively, it
might follow a fast direct route. The delay is very unpre-
dictable. Also the amount of time it takes a packet to
move across a network connection defines the Latency.
Latency and bandwidth are the two factors that deter-
mine a network connection speed. Latency and through-
put are two fundamental measures of network perform-
ance. Moreover, sometimes the routers might fail to de-
liver (drop) some packets (packet loss) if they arrive
when their buffers are already full. Some, none, or all of

S. NASRI353


the packets might be dropped, depending on the state of
the network, and it is impossible to determine what hap-
pened in advance. The receiving application must ask for
this information to be retransmitted, possibly causing
severe delays in the overall transmission [11].
3.3. QoS Modeling and Measurements
A traffic contract (SLA, Service Level Agreement) spe-
cifies the ability of a network or protocol to give guaran-
teed performance, throughput or latency bounds based on
mutually agreed measures, usually by prioritizing traffic.
A defined Quality of Service may be required for some
types of network real time traffic or multimedia applica-
tion [12-14]. We propose an approach of QoS-metric
based on QoS-parameter prioritization factors αi for one
application-service using the relation:

123
,,,,,, 1,,
mii
Qp pppFa pim (1)
We define k, αi, pi, and such as:

123
,,,,
m
Qp ppp
1) k ≥ 1: network efficiency coefficient ( in our case we
chose k = 1.1 for example).
2) αi: parameter prioritization factor, with:

1
1
m
i
i



 (2)
3) p
i: QoS performance parameter, pi should be nor-
malized pin
 
max min
max, min,
iii
ppp
i
p
a) For increasing parameters when data rate increases
min
max min
*
ii
in
ii
pp
pkp p

 (3)
b) For decreasing parameters when data rate increases
max
max min
*
ii
in
ii
pp
pkp p

 (4)
4) Then the QoS expression can be defined by:

2
1
,,,
m
imi
i
Qp ppp 

 

in
We use for this study the available network simulator NS.
This tool is becoming one of the most popular platforms
r size may drive
e NoC high performances. The main idea is to keep the
(5)
In this model we consider the packet loss parameter as
p1 and the EED parameter as p2 for FIFO and DWRR
scheduling techniques for 64 and 128 bytes of buffer size,
α1, α2 are arbitrarily fixed referring to the Equation (2).
4. QoS Behavior Simulation of Target
Architecture
for performances analysis in the network research com-
munity. Traffic is transferred over the network between
two IPs connected respectively to the 00 switch (source)
and the 33 switch destination. The packet size is 4 bytes
(32 bits) based on 8 bits/flits (4 flits per packet). The
maximum bandwidth link is fixed to 2GB/s. The purpose
of the study is to give a QoS measurements approach
according to the general network loading states and also
according to the interconnected IPs throughput [15].
4.1. Packets Loss and Buffer Size
As being discussed, a reasonable buffe
th
minimum buffer size avoiding dropped packets and EED.
We focus on percent flit losses due to buffer congestion
for a network loading. The routing approach is based on
the Wormhole Routing method. It reduces the store-
and-forward delay at each switch, and requires less
buffer size. Figures 2 and 3 show the relationship be-
tween percent dropped packet and available switch
buffer size. These figures show that the percentage of
dropped packets increases with application rate and de-
creases with the Buffer Size increase. We compare the
behavior of some parameters such as delay, dropped
packet and buffer size for the same architecture based on
the DWRR and FIFO scheduling. These figures testify
the capability of the DWRR to provide best results. In
Figure 2. Dropped packe t s for 64 by te s buffer size.
Figure 3. Dropped packets for 128 bytes buffer size.

S. NASRI 354
fact the percentage of dropped packets is significantly
less with DWRR compared to FIFO scheduling.
4.2. End to End Delay and Buffer Size
The EED is one of the most critical QoS metrics. Some
real-time applications bound up this value and require
specific hardware resources and particular management
approaches in the NoC switch. The best case is to pro-
vide the shortest constant EED or at least with the mini-
mum fluctuation. This can avoid synchronization be
twe p
heduler
ould improve service quality according to the flit re-
-
en communication processes. Figures 4 and 5 sum u
the EED when the switching buffer is managed with
DWRR and FIFO scheduling approach. The sc
sh
quirement, using priority queuing technique.
Figures 4 and 5 show that the EED average when
DWRR and FIFO scheduling technique are applied. It
decreases significantly with a buffer size value.
4.3. QoS Measurements
Referring to the proposed model the following figures
give the QoS measurements for 2 parameters: p1: packet
Figure 4. End to end delay according to the application rate
with DWRR and FIFO scheduling (64 byte s buffer size).
Figure 5. End-to-end delay according to the application rate
with DWRR and FIFO scheduling. (128 bytes buffer size).
loss and p2: EED, with prioritization factors: α1 = α2 =
0.5 and α1 = 0.2, α2 = 0.8.
Figures 6, 7 and 8 show the % QoS in relation with
the Buffer Size, the rate, the scheduling techniques and
prioritization factors. It appears that the % QoS increases
with the rate. The prioritization factors have also an im-
pact on the QoS values.
Figure 6. %QoS for 64bytes buffer size with prioritization
factors α1 = α2 = 0.5.
Figure 7. %QoS for 128 bytes buffer size with α1 = α2= 0.5.
Figure 8. %QoS for 128 buffer size with α1 = 0.2, α2 = 0.8.

S. NASRI 355
5. Conclusions and Perspectives
In this paper we have proposed a new QoS metric model
for Network on Chip based on the QoS paremeters for
one class of application. This model is a new approach of
QoS metric leading to quantify and measure a QoS value
in a network. We have focused our study on the sw
buffering requirements. In fact, we have showed that the
adequate buffer size in the switch drives to a better QoS
values. During NoC communication processes, QoS
metric can be affected by the switch buffering capacity
and its management approach. We have shown thath
DWRR is the best approach to manage packets schedul
i
ch, and then in
e network, can be ameliored by the adaptation of
itch
te
-
ng in the NoC switch (Figures 2, 3, 4 and 5). We think
that the QoS metric evaluation in the swit
th an
appropriate approach for packets queuing in the switch
buffer such as priority queuing (Figures 6, 7 and 8).
For this purpose, we are now working on the specifi-
cation of a new QoS model taking in consideration mul-
tiple applications with multiple QoS classes modeling.
The idea is to meet both network required performances
through the QoS metrics requirement, quantification and
measurments.
6. References
[1] International Sematech, “International Technology Road-
map for Semiconductors,” 2006.
http://public.itrs.net
[2] T. Bjerregaard and S. Mahadevan, “A Survey 0f Research
and Practice of Network on Chip,” ACM Computing Sur-
vey, Vol. 38, March 2006, pp. 1-51.
doi:10.1145/1132952.1132953
[3] J. Kim, D. Park, Ch. Nicopoulos, N. Vijaykrishnan and R.
Chita. Das, “Design and Analysis of an NoC Architecture
from Performance, Reliability and Energy Perspective,”
ACM symposium on Architecture for networking and
communications systems, Princeton, NJ, USA, 2005, pp.
173-182, ISBN: 1-59593-082-5.
Coenen, A. Radulescu, K. Goossens and G.
De Micheli, “A Methodology for Mapping Multiple
pe, Munich, Germany,
s on Chip,” Conference
ional Sympo-
Point to Point, Bus, and Net-
ware Resources Adap-
[4] S. Murali, M.
Use-Cases onto Networks on Chips,” Conference on de-
sign, automation and test in Euro
2006, pp. 118-123, ISBN:3-9810801-0-6.
[5] E. Rijpkema, K. Goossens and A. Radulescu, “Tradeoffs
in the Design of a Router With Both Guaranteed And
Best-Effort Services for Network
on design, automation and test in Europe (DATE’03),
March 2003, pp. 350-355.
[6] C. Grecu, A. Ivanov, P. Pande, A. Jantsch, E. Salminen,
U. Ogras and R. Marculescu, “Towards Open Net-
work-on-Chip Benchmarks,” First Internat
sium on Networks-on-Chip (NOCS’07), IEEE Computer
Society, May 2007, pp. 205-213.
[7] H. Gyu Lee, N. Chang, U. Y. Ogras and R. Marculescu,
“On-Chip Communication Architecture Exploration: A
Quantitative Evaluation of
work-on-Chip Approaches,” ACM transaction on Design
Automation of Electronic Systems, Vol. 12, No. 3, August
2007, pp. 1-20.
[8] A. Helali and S. Nasri, “Network on Chip Switch Sched-
uling Approach for QoS and Hard
tation,” International Journal of Computer Sciences and
Engineering Systems (IJCSES), Vol. 3, No. 1, 2009, pp.
29-35.
[9] T. Samak, E. Al-Shaer and H. Li, “QoS Policy Modeling
and Conflict Analysis,” IEEE Workshop on policy for
distributed systems and networks, 2008, pp. 19-26.
109/TWC.2010.061010.091803
[10] R. P. Liu, G. J. Sutton and I. B. Collings, “A New Queu-
ing Model for QoS Analysis of IEEE 802.11 DCF with
Finite Buffer and Load,” IEEE Transaction on Wireless
Communications, Vol. 9, No. 8, 2010, pp. 2664- 2675.
doi:10.1
appli-
Queuing Model
national sympo-
al conference on industrial
[11] V. X. Tran and H. Tsuji, “A Survey and Analysis on Se-
mantics in QoS for Web Services,” International Con-
ference on advanced information networking and
cations, October 2009, pp. 1-19.
[12] Y. Liu and H. He, “Grid Service Selection Using QoS
Model,” Third international conference on semantics,
knowledge and grid, 2007, pp. 576-577.
[13] J. Luo, L. Jiang, and C. He, “Finite
Analysis for Energy and QoS Tradeoff in Conten-
tion-Based Wireless Sensor Networks,” International
conference on communication (ICC), IEEE Communica-
tions Society, 2007, pp. 1-6.
[14] K. Kunavut and T. Sanguankotchakorn, “Multi-Con-
strained Path (MCP) QoS Routing in OLSR Based on
Multiple Additive QoS Metrics,” Inter
sium on communications and information technologies
(ISCIT-IEEE), 2010, pp. 226-231.
[15] Z. Zhou, W. Xu, D. T. Pham, C. Ji, “QoS Modeling and
Analysis for Manufacturing Networks: A Service Frame-
work,” 7th IEEE Internation
informatics (INDIN 2009), 2009, pp. 825-230.