Wimax Technical Report
Wimax Technical Report
Wimax Technical Report
The broadband wireless world is moving towards the adoption of WiMAX (the com-
mercial name of the IEEE 802.16 standard) as the standard for broadband wireless
Internet access. This will open up a very large market for industry and operators,
with a major impact on the way Internet access is conceived today. On the other
hand, the emergence of innovative multimedia broadband services is going to impose
severe Quality-of-Service (QoS) constraints on underlying network technologies. In
this work, after a brief review of the IEEE 802.16 standard, we intend to present
an in-depth discussion of its QoS support features. We point out the scheduling
algorithm as the critical point in QoS provisioning over such networks, and discuss
architectural and algorithmic solutions for an efficient support of multimedia flows.
Performance measurements obtained from an experimental test-bed are also pre-
sented with particular emphasis on the E-model calculation. The paper concludes
with a description of the key research challenges in the area, and provides a roadmap
for the research in the field.
1 Introduction
The IEEE 802.16 standard [1], promoted by the WiMAX (Worldwide Interoperabil-
ity for Microwave Access) forum (http://www.wimaxforum.org), will be the leading
technology for the wireless provisioning of broadband services in wide area net-
works. Such technology is going to have a deep impact on the way Internet access
is conceived, by providing an effective wireless solution for the last mile problem.
The market for conventional last mile solutions (e.g., cable, fiber etc.) presents
indeed high entrance barriers, and it is thus difficult for new operators to make
their way into the field. This is due to the extremely high impact of labor-intensive
tasks (i.e., digging up the streets, stringing cables etc.) that are required to put the
0
Core Network
SOHO
customer
Multi-tenant
customers
Base Station
will often allow only smaller channels (10 MHz or less) reducing the maximum
bandwidth. Even though 50 km distance is achievable under optimal conditions and
with a reduced data rate (a few Mb/s), the typical coverage will be around 5 km in
non-line-of-sight conditions and around 15 km with an external antenna in a line-of-
sight situation. Moreover, such a wide coverage makes it possible, and economically
viable to provide broadband connectivity in rural and remote areas, a market which
is usually not covered by traditional service providers.
In this paper, after a brief review of the standard fundamentals, we will provide
an in-depth overview and discussion on the QoS support provided by WiMAX tech-
nology. Particular attention will be devoted to scheduling algorithms for WiMAX
networks. We will survey the existing literature, and point out some common is-
sues involved in well-known technologies (e.g., wireless ATM), from which a system
designer can draw to design an efficient scheduler without starting from scratch.
Performance measurements obtained from an experimental test-bed are also pre-
sented with particular emphasis on the E-model calculation. The paper concludes
2
with an overview of the actual research challenges, pointing out and detailing the
most promising directions to pursue for research in this field.
In its first release in 2001, the 802.16 standard addressed applications for a fixed
scenarioin licensed frequency bands in the range between 10 and 66 GHz, where the
use of directional antennas are mandatory to obtain satisfactory performance figures.
In a metropolitan sub-area, however, line-of-sight operations cannot be ensured due
to the presence of obstacles, buildings, foliage etc. Hence, subsequent amendments
to the standard (802.16a and 802.16-2004) have extended the 802.16 air interface
to non-line-of-sight applications in licensed and unlicensed bands in the 2 − 11 GHz
frequency band. With the revision of IEEE standard document 802.16e, also some
mobility support will be provided. Revision 802.16f is intended to improve multi-hop
functionality, and 802.16g is supposed to deal with efficient handover and improved
QoS.
The transmission convergence sublayer operates on top of the PHY and provides
the necessary interface with the MAC. This layer is specifically responsible for the
transformation of variable-lenght MAC PDUs into fixed lenght PHY blocks [7].
4
The necessity to provide secure data transmissions has led to the native inclusion
of a privacy sub-layer, at the MAC level. Such protocol is responsible for encryp-
tion/decryption of the packet payload, according to the rules defined in the standard
[1].
Since IEEE 802.16 uses a wireless medium for communications, the main target
of the MAC layer is to manage the resources of the radio interface in an efficient
way, while ensuring that the QoS levels negotiated in the connection setup phase
are fulfilled. The 802.16 MAC protocol is connection-oriented and is based on a
centralized architecture. All traffic, including inherently connectionless traffic, is
mapped into a connection which is uniquely identified by a 16-bit address.
The common part sublayer is responsible for the segmentation and the reassembly
of MAC service data units (SDUs), the scheduling and the retransmission of MAC
PDUs. As such, it provides the basic MAC rules and signalling mechanisms for
system access, bandwidth allocation and connection maintenance. The core of the
protocol is bandwidth requests/grants management. A SS may request bandwidth,
by means of a MAC message, to indicate to the BS that it needs (additional) up-
stream bandwidth. Bandwidth is always requested on a per-connection basis to
allow the BS uplink scheduling algorithm (which is not specified in the standard) to
consider QoS-related issues in the bandwidth assignment process.
Regarding the way bandwidth is granted, the original 2001 standard encom-
passed two operational modes: Grant per Connection (GPC) and Grant per Sub-
scriber Station (GPSS). In the latest 2004 release, the term “grant” refers only
to the GPSS mode. In the GPC mode, the BS allocates bandwidth to individual
connections. This defines a purely centralized mechanism with all the intelligence
placed in the BS, while the SSs act as merely passive stations. On the other hand
the bandwidth, in the GPSS mode, is granted to each individual SS, which is then in
charge of allocating the available resources to the currently active flows. This defines
a semi-distributed approach to medium access control, in which some intelligence is
moved from the BS to the SSs. The reasons that led to the exclusion of GPC are
related to the fact that the BS scheduler may have scarce or outdated information on
the SSs queues status, possibly leading to suboptimal resource utilization. Further,
scalability issues may arise due to the complexity of maintaining entries for each
2. WIMAX TECHNOLOGY OVERVIEW 5
connection.
As depicted in Fig. 2, the MAC includes a convergence sublayer which provides
three main functionalities:
1. Classification. The CS associates the traffic coming from upper layer with an
apropriate Service Flow (SF) and Connection Identifier (CID).
2. Payload Header Suppression (PHS). The CS may provide payload header sup-
pression at the sending entity and reconstruction at the receiving entity.
The standard defines two different Convergence Sublayers for mapping services to
and from IEEE 802.16 MAC protocol. The ATM convergence sublayer is defined for
ATM traffic, while the packet convergence sublayer is specific for mapping packet-
oriented protocol suites, such as IPv4, IPv6, Ethernet and Virtual LAN. As regards
IP, the packets are classified and assigned to the MAC layer connections based on
a set of matching criteria, including the IP source and the destination addresses,
the IP protocol field, the Type-of-Service (TOS) or DiffServ Code Points (DSCP)
fields for IPv4, and the Traffic Class field for IPv6. However, these sets of matching
criteria are not in the standard and their implementation is left open to vendors.
Two main parameters are used in order to support service differentiation at the
higher layers: the Committed Information Rate (CIR) and the Maximum Informa-
tion Rate (MIR), inherited from other existing technologies [8, 9]. Both parameters
are set for a certain service class and regulate the entire aggregated downlink and
uplink flows of a given SS connection.
The CIR parameter for a WiMAX system is the bitrate that the network agrees to
accept from the user. In case of congestion, throughput reduction may occur below
the CIR: thus, the world “committed” is by no mean a guarantee that the CIR will
be met. A proper design of the user network, anyhow, should make this event quite
6
rare.1
Flows exceeding the CIR are vulnerable to packet discarding policies at the operator
need: if the WiMAX network is congested, the BS will typically discard frames on
connections exceeding the CIR before frames on connections that are within their
CIR. Thus, the CIR provides a crude method for being fair when allocating limited
capacity.
The second parameter, the MIR, regulates the maximum allowed peak rate of a
connection. If the transmission rate exceeds the MIR, all the MAC frames violating
the MIR will be discarded automatically; usually, the exact details on how the BS
discarding policies is proprietary to the hardware vendor.
MIR and CIR are specified for each SS according to the negotiated Service Level
Agreement (SLA); the compliance to the negotiated SLA is assessed over a reference
window, called Committed Time (CT). In what follows we assume that n SSs make
MIR and CIR requests to the BS. We let Rmax the maximum traffic rate available
at the WiMAX Downlink Air Interface, and denote CIRk and MIRk the request of
the k-th SS2 , where 0 ≤ CIRk ≤ MIRk ≤ Rmax .
The BS dynamically allocates the BE Service Rate RBE (bit/s) and the Real Time
(RT) Service Rate RRT (bit/s) with a cumulative upper bound of Rmax , making
sure that the RT service traffic has a higher priority than the BE service traffic:
RRT + RBE ≤ Rmax . The residual rate is allocated to RBE : the usual assumption is
that the BE flows are TCP friendly, which is an assumption verified in our experi-
ments. Let Ntot be the total number of downstream service flows consisting of NVoIP
VoIP flows and NTCP TCP persistent connection, so that Ntot = NTCP + NVoIP .
Let RTCP (m) be the service rate that the BS can provide to the m-th TCP ser-
P TCP
vice flow, the aggregated BE service rate is RBE = Nm=1 RTCP (m); similarly, if
RVoIP (m) is the service rate that the BS provides to the m-th VoIP service flow,
P VoIP
the aggregated RT service rate becomes: RRT = N m=1 RVoIP (m). The Alvarion
equipment used in the testbed provides resource allocation mechanisms responding
1
If the customer has negotiated a Service Level Agreement with the service provider, the service
provider should pay a penalty for missing a CIR commitment.
2
In Alvarion the MIR for real-time traffic is not defined
2. WIMAX TECHNOLOGY OVERVIEW 7
to three cases.
In the first case, the downlink bandwidth is over provisioned, meaning that the ag-
gregated traffic service rate for the WiMAX network is deterministically lower than
PNTCP PNVoIP
Rmax , i.e. m=1 MIR(m) + n=1 MIR(n) ≤ Rmax , and no congestion occurs:
the allocation in this case is fairly simple and the BS sets RVoIP (n) = MIR(n) and
RTCP (m) = MIR(m).
The opposite case occurs when the aggregate of the CIR requested by VoIP sub-
PNV oIP
scribers exceeds Rmax , i.e. n=1 CIR(n) > Rmax : then the BS sets RVoIP (n) =
Rmax
NV oIP
and RTCP (m) = 0 for every SS n = 1, 2, . . . , n.
The remaining case is such that
X
N TCP NX
VoIP
This is the case when the BS guarantees the minimum service rate for the VoIP
traffic and can reallocate the remaining bandwidth to the BE services, namely
As described above, the data packets entering the IEEE 802.16 network are mapped
into a connection and a service flow based on a set of matching criteria. These
classified data packets are then associated with a particular QoS level, based on the
QoS parameters of the service flow they belong to. The QoS may be guaranteed by
8
shaping, policing, and/or proritizing the data packets at both the SS and BS ends.
The BS allocates upstream bandwidth for a particular upstream service flow based
on the parameters and service specifications of the corresponding service scheduling
class negotiated during connection setup. The IEEE 802.16 standard defines four
QoS service classes: Unsolicited Grant Service (UGS), Real-Time Polling Service
(rtPS), Non-Real Time Polling Service (nrtPS) and Best Effort (BE) [10][7].
The four classes are characterized as follows.
• The UGS service is defined to support constant bit rate (CBR) traffic, such as
audio streaming without silence suppression. Unsolicited grants allow SSs to
transmit their PDUs without requesting bandwidth for each frame. The BS
provides fixed-size data grants at periodic intervals to the UGS flows. Since
the bandwidth is allocated without request contention, the UGS provides hard
guarantees in terms of both bandwidth and access delay. The QoS parameters
defined for this service class are the grant size to be allocated, the nominal
interval length between successive grants and the tolerated grant jitter, defined
as the maximum tolerated variance of packet access delay.
• In the case of Variable Bit Rate (VBR) video traffic, such as MPEG streams,
the bandwidth requirements for the UGS grant interval cannot be determined
at connection setup time. As a result, peak stream bit rate-based CBR allo-
cation would lead to severe network underutilization, whereas the average bit
rate CBR allocation can result in unacceptable packet delay and jitter. The
rtPS service has been introduced to accomodate such flows. For this service,
indeed, the BS provides periodic transmission opportunities by means of a
basic polling mechanism. The SS can exploit these opportunities to ask for
bandwidth grants, so that the bandwidth request can be ensured to arrive
at the BS within a given guaranteed interval. The QoS parameters relevant
to this class of services are the nominal polling interval between successive
transmission opportunities and the tolerated poll jitter.
• The non real-time polling services (nrtPS) is similar in nature to rtPS but
it differs in that the polling interval is not guaranteed but may depend on
the network traffic load. This fits bandwidth-demanding non-real time service
3. QOS SCHEDULING IN WIMAX NETWORKS 9
flows with a variable packet size, such as large files transfers. In comparison
with rtPS, the nrtPS flows has fewer polling opportunities during network
congestion, while the rtPS flows are polled at regular intervals, regardless
of the network load. In heavy traffic conditions, the BS can not guarantee
periodic unicast requests to nrtPS flows, so that the SS would also need to use
contention and piggybacking to send requests to the BS uplink scheduler.
• For Best Effort (BE) traffic, no periodic unicast requests are scheduled by the
BS. Hence, no guarantees in terms of throughput or packet delay can be given.
The BE class has been introduced to provide an efficient resource utilization
for low-priority elastic traffic, such as telnet or HTTP.
While these services provide the basics for supporting QoS guarantees, the “real”
core, i.e., traffic scheduling, policing, shaping and admission control mechanisms, is
not specified by the standard. In the next section, we will present and review some
possible QoS architectures for WiMAX-based PMP networks.
In order to offer an efficient QoS support to the end user, a WiMAX equipment
vendor needs to design and implement a set of protocol components that are left
open by the standard. These include traffic policing, traffic shaping, connection
admission control and packet scheduling.
Due to the highly varying nature of multimedia flows, traffic shaping and traffic
policing are required by the SS, in order to ensure an efficient and fair utilization of
network resources. At connection setup, the application requests network resources
according to its characteristics and to the required level of service guarantees. A
traffic shaper is necessary to ensure that the traffic generated actually conforms to
the pre-negotiated traffic specification. However, traffic shaping may not guarantee
such conformance between the influx traffic and service requirements. This is dealt
with by a traffic policer, which compares the conformance of the user data traffic with
the QoS attributes of the corresponding service and takes corresponding actions, e.g.,
it rejects or penalizes non conformance flows.
10
QoS profiles for SS are usually detailed in terms of Committed Information Rate
(CIR) and Maximum Information Rate (MIR) for the various QoS classes [8]. The
CIR (defined for nrtPS and rtPS traffic) is equal to the information transfer rate that
the WiMAX system is committed to carry out under normal conditions. While the
MIR (defined for nrtPS and BE QoS types) is the maximum information rate that
the system will allow for the connection. Both this QoS parameters are averaged
over a given interval time.
In order to guarantee that the newly admitted traffic does not result in net-
work overload or service degradation for existing traffic, a (centralized) connection
admission control scheme also has to be provided.
Even though all the aforementioned components are necessary in order to provide
an efficient level of QoS support, the core of such a task resides in the scheduling
algorithm. An efficient scheduling algorithm is the essential conditio sine qua non
for the provision of QoS guarantees, and it plays an essential role in determining the
network performance. Besides, a traffic shaper, policer and connection admission
control mechanisms are tightly coupled with the scheduler employed. Therefore, the
rest of this section is devoted to such an issue.
Although the scheduling is not specified in the standard, system designers can
exploit the existing rich literature about scheduling in wireless ATM [11], from which
WiMAX has inherited many features. If this allows one not to start from scratch,
existing schemes need to be adapted to match the peculiar features (e.g., traffic
classes, frame structure) of the IEEE 802.16 standard.
As an example, the GPSS scheduling mode can be seen as an outcome of the
research carried out on hierarchical scheduling [12]. This is rooted in the necessity
of limiting the MAC exchange overhead by letting the BS handle all connections of
each SS as an aggregated flow. As explained in the previous section, according to
the standard, the SSs request bandwidth on per connection-basis; however, the BS
grants bandwidth to each individual SS, so that the resources are allocated to the
aggregation of active flows at each SS. Each SS is then in charge of allocating the
granted bandwidth to the active flows, which can be done in an efficient way since
the SS has complete knowledge of its queues status. This, however, requires the
introduction of a scheduler at each SS, enhancing the complexity (and consequently
3. QOS SCHEDULING IN WIMAX NETWORKS 11
status to the BS for the uplink. This has a twofold effect. On the one hand, it
increases the signalling overhead, while, on the other, it provides the BS with infor-
mation that may be not up-to-date (e.g., due to contention delays etc.). In downlink,
the scheduler has complete knowledge of the queue status, and, thus, may use some
classical scheduling schemes, such as Weighted Round Robin (WRR), Weighted Fair
Queueing (WFQ) etc. [11]. Priority oriented fairness features are also important in
providing differentiated services in WiMAX networks. Through priority, different
traffic flows can be treated almost as isolated while sharing the same radio resource.
However, due to the nature of WiMAX TDD systems, the BS scheduler is non-
work-conserving, since the output link can be idle even if there are packets waiting
in some queues. Indeed, after downlink flows are served in their devoted subframe,
no additional downlink flows can be served till the end of the subsequent uplink
subframe.
Scheduling uplink flows is more complex because the input queues are located
in the SSs and are hence separated from the BS. The UL connections work on a
request/grant basis. Using bandwidth requests, the uplink packet scheduling may
retrieve the status of the queues and the bandwidth parameters. The literature
is not rich in terms of QoS scheduling schemes specifically designed for WiMAX
networks. In the following, we will briefly describe the most relevant works that
address such a topic, to the best of authors knowledge.
In [13], the authors present a QoS architecture for IEEE 802.16 based on pri-
ority scheduling and dynamic bandwidth allocation. In particular, they propose a
scheduling process divided into two parts. The first one, executed by the uplink
scheduler inside the BS, is performed in order to grant resources to the SSs in re-
sponse to bandwidth requests. This is done by means of a classical WRR [14]. At
each subscriber station, bandwidth assignments are computed by starting from the
highest priority class (i.e., UGS flows) and then going down to rtPS, nrtPS and BE.
In this way, a strict priority among service classes is guaranteed. The scheduling
schemes employed for the various classes are different. A classical WFQ [15] is used
for UGS and rtPS, whereas a simpler WRR is used for nrtPS service class. Best
Effort traffic is served through a simple FIFO policy. Inside each class, a Multiclass
Priority Fair Queuing (MPFQ) [12] scheduler, with some slight adjustments, is em-
3. QOS SCHEDULING IN WIMAX NETWORKS 13
ployed. The four service classes have a strict priority and each of the priority classes
have a packet scheduler with its own policy. A packet ready to be transmitted is
chosen from the highest priority class. If connections belonging to lower priority
class want to send a packet, they have to wait the end of the transmission in a
higher class.
In order to provide a lower delay bound for the higher priorities (UGS and rtPS),
a Wireless Fair Queuing [15] policy is used. Weighted Round Robin (WRR) [14]
scheduling is adopted for the nrtPS classes since the delay requirements are not so
tight. Finally a FIFO scheduling is implemented for the BE traffic.
By means of this prioritized approach (which resembles somehow Multiclass Priority
Fair Queueing [12]), the proposed architecture is able to guarantee a good perfor-
mance level to UGS and rtPS classes, to the detriment of lower priority traffic (i.e.,
nrtPS and BE flows).
The other work in the literature [16] presents an approach based on a fully
centralized scheduling (GPC-like) scheme, where a global QoS agent collects all
the necessary information on traffic flows, and takes decisions on traffic admission,
scheduling, and resource allocation. Based on the complete global knowledge of the
system, the deterministic quality of service levels can be guaranteed. Their scheme
shares some common features with [13] concerning how the priorities are treated.
And even in [16] strict priority among different service classes is provided. In terms
of the scheduling discipline used for the various classes, both Earliest Deadline First
(EDF) [17] and WFQ are used. Still, the strict pririty discipline allow to redistribute
bandwidth among its active connections from highest to lowest: UGS, rtPS, nrtPS
and BE. In particular, EDF is used for rtPS flows, that demands the insertion of
an additional module in the uplink scheduling for computing the packet deadline.
In general, EDF is able to outperform standard WFQ in terms of performance, but
the complexity and the cost of this additional module is the price to pay for it.
As the authors take into consideration a completely centralized scheme in [16], this
cost is negligible. However, this may become a limiting factor in the case of a semi-
distributed (i.e., GPSS) scheduling mechanism. WFQ is used to share the bandwidth
among the nrtPS schemes, while a simple FIFO mechanisms is encompassed for
elastic (e.g., BE) traffic.
14
Finally, in [10] the authors have assesed, via simulation, the performance of an
IEEE 802.16 system using the class of latency-rate scheduling alghoritms where a
minimum reserved rate is the basic QoS parameter negotiated by a connection within
a scheduling service. Specifically, within this class, they selected defict round robin
(DRR) as the downlink scheduler to be implemented in the BS, since it combines
the ability to provide fair queuing in the presence of variable lenght packets with
the simplicity of implementation. In particular, DRR requires a minimum rate to be
reserved for each packet flow being scheduled. Therefore, althought not required by
the IEEE 802.16 standard, BE connections should be guaranteed a minimum rate.
This fact can be exploited in order to both avoid BE traffic starvation in overloaded
scenarios, and let BE traffic take advantage of the excess bandwidth which is not
reserved for the other scheduling services. On the other hand, DRR assumes that
the size of the head-of-line packet is known at each packet queue; thus, it cannot
be used by the BS to schedule transmissions in the uplink direction. In fact, with
regard to the uplink direction, the BS is only able to estimate the overall amount of
backlog of each connection, but not the size of each backlogged packet. Therefore,
the authors selected WRR as the uplink scheduler. Like DRR, WRR belongs to the
class of rate-latency scheduling algorithms. At last, DRR is implemented in the SS
scheduler, because the SS knows the sizes of the head-of-line packets of its queues.
Internet
Router
support Voice over IP (VoIP) flows mapped on different QoS service classes with
different background traffic mixes. Voice quality is evaluated by using the com-
putational scheme provided by the E-Model [20]. This scheme takes into account
different parameters affecting the quality of a conversation and can be related to the
Mean Opinion Score (MOS). The MOS is the widely accepted standard for speech
quality rating, used to measure the speech quality in a scale from 1 (poor quality)
up 5 (excellent). Now, the Packet E-Model is introduced.
5 Packet-E-Model
which provides an objective method to evaluate speech quality in VoIP systems (for
a thorough description see [22, 23]). The outcome of an E-Model evaluation is called
R-factor (R). The R-factor is a numerical measure of voice quality, ranging from 0
to 100. The reference values of the R-factor are categorized as shown in Table 1.
In the E-Model several different parameters affecting the quality of a conversation
are taken into account. The main assumption is that various impairments at the
psychological scale have an additive behavior (dB-like behavior),
R = R0 − Is − Id − Ie + A. (3)
Table 3: Equipment impairment factor for the G.729.2 and G.723.1 codecs
Parameter G.729.2 G.723.1
Ieopt 10 15
C1 47.82 90
C2 0.18 0.05
where H(x) is the step function and Ief is the equipment impairment (non-linear
codecs and packet losses). Ief is calculated as4
Clearly, Ieopt , C1 and C2 are codec specific parameters: Table 3 reports the values
for the two codecs employed in our tests [24].
6 Testbed Configuration
Our testbed targets a residential broadband access, where the system operates in
the 2−11 GHz band. The experimental data has been collected exploiting a 4-nodes
wireless testbed deployed in a rural environment, implementing a PMP structure, as
sketched in Fig. 4. The BS is equipped with a sectorial antennas with a gain of 14
dBi in the vertical plane covering all the 3 SSs. The default maximum output power
at antenna port is 36 dBm for both the BS and the SS. The distance between the
BS and SS1, SS2 and SS3 is 8.4 km, 8.5 km and 13.7 km, respectively. The average
signal-to-noise ratio is above 30 dB, thus enabling the higher modulation, i.e. 64
QAM, for each connection. The SSs work in line-of-sight conditions under FDD half-
duplex. All nodes run Linux with a 2.4 kernel. The measurements are performed
exploiting an Alvarion BreezeMAX platform operating in the 3.5 GHz licensed band
and using a 3.5 MHz wide channel in FDD mode. Each node is attached through
an Ethernet connection to the WiMAX equipment.
4
The calculation is accurate up to loss rate = 0.1, for higher values it may prove optimistic.
18
• CIR. The CIR is defined for rtPS and nrtPS traffic only. The range is from 0
to 12 Mbps that is the maximum (MAC) throughput of Alvarion BreezeMAX
equipment.
• MIR. The MIR is defined for nrtPS and BE QoS types and the rate is averaged
as in the case of the CIR.
• CT. The CT defines the time window over which the information rate is aver-
aged to ensure compliance with the CIR or MIR parameter.
The CT allowed values are reported in Tab 4. The IP’s DSCP [25] field is ex-
ploited in order to enforce a certain QoS class service. Traffic flows belonging to
different service categories are tagged using the iptables software [26]. During our
measurements, all SSs share the same QoS, as summarized in Tab. 5.
and inject different traffic patterns over TCP and/or UDP sockets. Traffic is then
collected at the receiver side where suitable tools are available for analysis. In our
settings, we assumed that several concurrent VoIP flows use a SS as their gateway
towards a peer terminal (this would be the typical case of several voice stations
multiplexed at a VoIP gateway). We measured the performances of the uplink and
the downlink separately, thus neglecting interference effects. VoIP codecs feed
RTP packet flows and three commonly used codecs have been considered (G.711,
G.729.2 and G.723.1), whose parameters are reported in Tab. 6. Also, the considered
scenario was homogeneous and the background data traffic (in our case persistent
TCP connections) was modeled considering a TCP socket working in saturation
regime, according to the parameters reported in Tab. 7. rtPS services are used for
VoIP connections, while TCP-controlled traffic is mapped in the BE class. The
mapping of CBR sources into the rtPS class made much easier trace the behavior
of the system, since the actual scheduling policies were unknown on our side.
In order to collect reliable measure of delays, before each experiment we synchronized
each node with a common reference using NTP [28]. All SSs sustain the same
traffic, consisting in a given number of increasing VoIP session plus one persistent
TCP connection (aimed at modeling background traffic). All measurements were
performed over 5 minutes intervals; results are averaged over 10 runs. In the next
section we report the performance obtained over the test-bed described above.
4
x 10
8
G.723.1
G.729.2
7 G.711
6
Session throughput (bit)
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Number of concurrent VoIP calls
7 Performance Measurements
In the first set of measurements, we determined the voice capacity, i.e. the maximum
number of sustained VoIP calls with high quality (70 < R < 80) and related param-
eters: VoIP throughput, delay and packet loss. Here, we report only the downlink
behavior, since we found that the downlink was actually the bottleneck.
We first measured the average throughput of VoIP calls, at the increase of the num-
ber of VoIP flows. It turned out that the performance for the G.711 codec are far too
low to be acceptable and a SS could not support more than 2 (high-quality) calls, as
depicted in Fig. 5. Hence, in the following we reported the comparison for the two re-
maining codecs. Fig. 6 and Fig. 7 depict the results for the delay and the packet loss,
respectively. The delay, in particular, saturates at 300 ms, whereas, after the satu-
ration point, packet loss increases almost linearly. The G.723.1 codec outperforms
clearly G.729.2; such a difference is due to the higher G.729.2 packet generation
rate, coupled to the large overhead of packet headers of the RTP/UDP/IP/MAC
protocol stack (≃ 45% for the G.729.2). Such effect is very well known in VoIP over
WLANs [29] or, more in general, for bandwidth-limited connections [23]. In prac-
7. PERFORMANCE MEASUREMENTS 21
G.729.2
0.3
0.25
Average delay (s)
0.2
0.15
0.1
0.05
G.723.1
8 9 10 11 12 13 14 15 16 17 18 19
Number of concurrent VoIP calls
Figure 6: Average delays versus an increasing number of concurrent VoIP flows per
SS for different codecs; minimum value and maximum value delimiters are superim-
posed.
G.723.1
45 G.729.2
40
35
Average packet loss (%)
30
25
20
15
10
0
9 10 11 12 13 14 15 16 17 18 19
Number of concurrent VoIP calls
Figure 7: Packet loss rate of VoIP flows per SS using different codecs.
22
85
G.723.1
G.729.2
80
75
Average R−factor
70
65
60
8 9 10 11 12 13 14 15 16 17 18 19
Number of concurrent VoIP calls
Figure 8: Average R-factor versus number of concurrent VoIP flows per SS using
different codecs.
tice, it is convenient to employ larger speech trunks per packet and consequently
larger inter-packet generation intervals.
Finally, Fig. 8 provides a comprehensive picture in terms of the R-Factor. We notice
that there exist roughly three regions: in the leftmost region, G.729.2 provides a
fairly good quality, but as soon as the network starts saturating around 10 calls,
G.723.1 obtains much better performance. When the network is overloaded, on the
other hand, we notice that the G.723.1 codec experiences a larger degradation than
the G.729.2. In the end, we could assess that with G.723.1, the system under exam
can support up to 17 VoIP calls per SS with a high quality. Conversely, the use of
a G.729.2 codec reduces voice capacity to 10.
In order to determine the voice capacity of the system, we restricted our focus
on the downlink, because it poses the most stringent constraints; in this section
we justify this statement showing the outcomes of the uplink and downlink. As
reported in Fig. 9, at the voice capacity, the quality of the perceived speech is better
for the uplink, irrespective of the index of the SS VoIP flow considered, and of the
7. PERFORMANCE MEASUREMENTS 23
78
76
74
72
R−factor
70
68
66
64
DL − G.729.2
62 UL − G.729.2
DL − G.723.1
UL − G.723.1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Number of VoIP calls
Figure 9: Uplink and downlink R-factor versus the SS VoIP session index, using 11
and 17 concurrent calls with the G.729.2 and G.723.1 codecs, respectively.
codec considered. Also, the packet loss was always slightly better for the the uplink
compared to the downlink, for this case and the following ones.
In order to have a better understanding of the system behavior, we sampled the
first order probability density function of the packet delay, both for the uplink and
the downlink in some critical cases, i.e. for a number of sessions around the voice
capacity. In particular, Fig. 10, Fig. 11 represent the delay pdf for downlink VoIP
flows. Even though the scheduling policy is undisclosed, it is apparent that it is not
simply the average delay to degrade a the increase of the offered VoIP traffic, but
the whole delay distribution is shifted around higher delay values.
This proves that the BS operates a very strict threshold control policy: in case
a SS exceeds a certain threshold above the CIR, the system basically penalizes
any violating SS. In fact, only for 17 G.723.1 VoIP calls the excess above the CIR
appears evenly redistributed over the interval, the rationale being that in such case
the smaller throughput of the codec might bring oscillations above and below the
limit. At the SS side, this strict BS policy indeed suggests to employ suitable
admission control for outgoing and incoming VoIP flows, in order not to incur into
major service degradation.
24
60
16 flows
17 flows
18 flows
50
Probability density function
40
30
20
10
0
0.05 0.1 0.15 0.2 0.25 0.3
Delay (s)
60
10 flows
11 flows
50
Probability density function
40
30
20
10
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Delay (s)
35
10 flows
11 flows
30
Probability density function
25
20
15
10
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Delay (s)
We repeated the same measurements in the case of the uplink, and, as reported
in Fig. 13 and Fig. 12, the results are similar. As emerged from the R-factor mea-
surements, the uplink performs better than the downlink and, in fact, the delay
distribution of the uplink around the VoIP capacity appears in all cases centered at
lower values compared to the downlink. We remark that the uplink measurements
contradict the simulation results obtained in [10], where larger delays in the uplink,
compared to the downlink, were ascribed to the bandwidth request mechanism and
to the PHY overhead. In the case at hand, the uplink delay due to bandwidth
request did not prove significant, and we ascribe this fact to the activation of the
piggybacking mechanisms for bandwidth reservation provided by WiMAX.
8 Research Challenges
Though WiMAX is the most promising technology for enabling BWA systems to be
widely deployed, many issues need to be addressed in order to make it effectively
support the requirements and constraints of end-users’ multimedia flows. In order
to do so, according to the discussion mentioned previously, an efficient QoS-enabled
scheduling algorithm has to be designed and implemented. In this section, we point
26
40
16 flows
17 flows
35 18 flows
30
Probability density function
25
20
15
10
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Delay (s)
out and briefly describe the most promising, as well as challenging, directions in
such a field, by outlining a research roadmap for QoS provisioning in WiMAX net-
works. As we considered the scheduling algorithm in isolation in the last section, we
shall now present cross-layer approaches, in which performance improvements are
obtained by making an appropriate use of information which comes from the lower
and/or upper layers.
QoS support. For example, improving the effect of non real-time traffic (i.e.,
nrtPS and BE traffic) would free some additional resources to higher priority
traffic. In this way, opportunistic scheduling schemes may actually help to
increase the QoS capabilities of WiMAX networks. Moreover in this case, novel
scheduling schemes are required in order to exploit multiuser diversity while
providing QoS guarantees to the active traffic flows at the same time. It may be
interesting to note that multiple antenna systems can actually be used to build
up multiuser diversity by means of random beamforming mechanisms (usually
referred to in the literature as “dumb” antennas [32]). While this direction is
somehow orthogonal in nature to the one (based on “smart antennas”) outlined
above, it could be worth investigating whether these two techniques may be
implemented to coexist (for example, in a time-sharing fashion) in order to
obtain the advantages of both approaches.
References
[1] IEEE Standard for Local and Metropolitan Area Networks Part 16: Air Inter-
face for Fixed Broadband Wireless Access Systems, IEEE Std., Rev. IEEE Std
802.16-2004, 2004.
[2] Nokia, “Broadband Media Services,” 2001, White Paper. [Online]. Available:
http://www.nokia.com/downloads/solutions/operators/broadband media services tutorial net
tential,” IEEE Communications Magazine, vol. 43, no. 2, pp. 129–136, February
2005.
[8] N. J. Muller and R. P. Davidson, The Guide to Frame Relay & Fast Packet
Networking, 1991.
[12] J.Moorman and J.Lockwood, “Multiclass priority fair queuing for hybrid
wired/wireless quality of service support,” in Proceedings of Second ACM Inter-
national Workshop on Wireless Mobile Multimedia, WOWMOM, August 1999.
[13] G. Chu, D. Wang, and S. Mei, “A QoS architecture for the MAC protocol of
IEEE 802.16 BWA system,” IEEE Conference on Communications, Circuits
and Systems and West Sino Expositions, vol. 1, pp. 435–439, June 2002.
[14] K. Mezger, D. W. Petr, and T. G. Kelley, “Weighted Fair Queuing vs. Weighted
Round Robin: A Comparative Analysis.” in Proceedings of IEEE Wichita Con-
ference on Communications, Networks and Signal Processing, 1994.
[16] K. Wongthavarawat and A. Ganz, “Packet scheduling for QoS support in IEEE
802.16 broadband wireless access systems,” International Journal of Commu-
nication Systems, vol. 16, pp. 81–96, 2003.
[21] ITU-T, “The E-model, a computational model for use in transmission plan-
ning,” Telecommunication Standardization Sector of ITU, Geneva, Switzerland,
Recommendation G.107, Dec. 1998.
[23] C. Hoene, H. Karl, and A. Wolisz, “A perceptual quality model intended for
adaptive voip applications: Research articles,” Int. J. Commun. Syst., vol. 19,
no. 3, pp. 299–316, 2006.
[25] “Definition of the differentiated services field (ds field) in the ipv4 and ipv6
headers.” [Online]. Available: http://www.ietf.org/rfc/rfc2474.txt
[28] “Simple Network Time Protocol (SNTP) Version 4.” [Online]. Available:
http://www.apps.ietf.org/rfc/rfc2030.html