The Road toward Internet Media

Voice-Quality Monitoring and Control for VoIP Voice over IP offers important opportunities for the telecommunications market to deploy more advanced services, but it must overcome many obstacles. Users expect toll-quality voice, which calls for end-to-end quality of service (QoS) — a challenge for IP service providers. To make VoIP attractive to end users, the only feasible and directly implementable alternative is to deploy an efficient mechanism within the endpoints.To that end, the authors propose the scalable, modular, Call Quality Monitoring and Control framework for maintaining voice quality at acceptable levels over networks that don’t offer QoS guarantees.

U

nlike past decades, during which almost all telecommunications network traffic was voice, today’s telecommunications market is driven largely by IP-oriented applications and technologies. The Internet’s proliferation has brought the volume of data traffic close to that of voice traffic. As a result, a shift is occurring toward technologies that were initially designed to serve merely data traffic but now extend to other services such as voice, video, or multimedia.1 Coupled with the need for interoperability between private and public networks, this move has driven telecom operators to adopt IP to unify traffic types. Such a vast change has its problems, however. Users are accustomed to the quality of service (QoS) they’ve enjoyed for years with the public switched telephone network (PSTN). Some are reluctant to use networks and systems based on voice over

IEEE INTERNET COMPUTING

IP (VoIP). Such reluctance is reasonable given that a best-effort network such as the Internet can’t guarantee when, or how much, data will get delivered — a fact that affects perceived service quality. With VoIP, quality includes both callestablishment parameters (such as service availability and call-setup time) and voice quality (VQ), which can suffer due to IP network impairments. Using prioritization and reservation frameworks or protocols can minimize these impairments’ impact on real-time applications. To deploy them, however, service providers must first agree on a common policy and then cooperate with each other in implementing it. This solution isn’t currently feasible because providers enforce their own policies in their networks. Consequently, endto-end QoS depends heavily on the strategies and control mechanisms implemented at the endpoints. Our endpoint-

1089-7801/05/$20.00 © 2005 IEEE

Published by the IEEE Computer Society

Michael Manousos, Spyros Apostolacos, Ioannis Grammatikakis, and Dimitrios Mexis inAccess Networks Dimitrios Kagklis and Efstathios Sykas National Technical University of Athens

JULY • AUGUST 2005

35

The Road toward Internet Media

based architecture — named Call Quality Monitoring and Control (CQMC) — can alleviate the endto-end QoS deterioration that comes from existing best-effort IP networks. We’ve implemented CQMC on a platform to show the system’s efficiency.

VQ Degradation Factors Many factors can affect voice transmitted through an IP network, including packet loss, delay, jitter, and voice encoding due to limited bandwidth. VoIP applications use various well-known mechanisms to combat packet loss due to networkrouter congestion, router or link failures, and bit errors during transmission. Such mechanisms are either receiver-based (such as silence substitution, noise substitution, packet repetition or interpolation, time-scale modification, and decoder-based loss concealment) or sender-based (such as packet retransmission, frame interleaving, and forwarderror correction [FEC]).2

Many factors can affect voice transmitted through an IP network, including packet loss, delay, jitter, and voice encoding. Delay in VoIP occurs due to voice processing (encoding, decoding, and lookahead — the large buffering of voice samples that certain codecs employ to improve compression ratio and compressed voice quality), voice-data packetization, packet transmission, and receiver playout delays (the buffering of voice frames arriving at the receiver to eliminate network jitter). Endpoints can control all these factors, except transmission delay, to minimize their effects on conversation quality. Jitter is the variation of the delay that consecutive packets experience due to variable waiting time inside router queues or to following different paths. VoIP applications can’t tolerate jitter, which can be eliminated with static or adaptive playout buffers.3 Finally, voice codecs employed to reduce bandwidth induce distortion, which end users perceive as VQ degradation. Unfortunately, we can’t treat these impairments separately. Optimization with respect to a single parameter has detrimental effects on the others. For example, frame interleaving and FEC mecha-

36

JULY • AUGUST 2005

www.computer.org/internet/

nisms decrease the effects of loss but increase the end-to-end delay, whereas a large playout buffer removes jitter but increases delay and loss.

VQ Control and Monitoring For these reasons, we must pursue a joint optimization strategy by maximizing VQ as a function of all variables that represent network impairments and operational system parameters. Determining this multivariate function’s optimum outcome results in extreme complexity; given the limited computational resources available in real-time, real-world systems, we must reduce the number of variables to achieve useful results.4,5 Our solution is to use realtime monitoring and adaptation of an extensive set of system operational parameters to track time-variant network conditions while maintaining VQ at the highest possible level. To successfully implement this solution, we must use a VQ-monitoring mechanism to drive the adaptive mechanism in real time. In the past, VQ evaluation was subjective and required an audience and listening tests.6 Such an approach is inappropriate for quality control requiring realtime monitoring. Objective VQ monitoring, whether active or passive, has recently gained ground among IP telephony service providers (ITSP). In active monitoring, a network analyzer injects traffic patterns that resemble a VoIP application into the network; the analyzer then observes the overall VQ by comparing the impaired voice with the original voice sample using a perceptual model.7 Although this scheme can provide useful input for optimization and network dimensioning, it uses network resources, provides non-real-time results, and can’t concretely determine the causes of degradation. A passive monitoring scheme, on the other hand, can operate in real time, and lets VoIP applications take corrective action when QoS is unacceptable. The E-model,8 as standardized by the ITU-T, uses a computational method to characterize conversation quality from mouth to ear as the receiving-side user perceives it, both as listener and talker. It then rates VQ using the transmissionrating factor, R, which ranges from zero to 100. A detailed description of its application to IP telephony systems is available elsewhere.9

Monitoring and Control Mechanism To satisfy the need for VQ control at best-effort networks’ endpoints adaptively and with low com-

IEEE INTERNET COMPUTING

Voice-Quality Monitoring

Call quality data provided for real-time monitoring

plexity by monitoring VQ in real time, we’ve implemented the CQMC mechanism. Its design fulfills two common requirements for VoIP gateways and IP private branch exchanges (IP-PBXs) that handle numerous IP calls. First, its functionality doesn’t depend on the VoIP layer implementation — the system supports both H.323 and the Session Initiation Protocol (SIP) to establish, control, and tear down IP calls; a media transport protocol (the Real-Time Transfer Protocol [RTP]/Real-Time Control Protocol [RTCP] stack); a playout buffer; and voice codecs (G.711, G.729, and GSM06.10). Second, its operation is scalable to a high number of concurrent calls, limited only by the VoIP gateway’s media-processing capabilities. As part of an IP-PBX, we segmented CQMC’s functionality into two parts: agents, which are VoIP layer implementation-aware parts, and the core, which is the main processing unit (see Figure 1).

Real- time data processing Monitoring and control

Data storage and retrieval

CQMC API CQMC core

IP-PBX switching core

Agents A CQMC agent collects raw data for both call directions, so that the core can perform callquality monitoring and control. For the ingress direction, agents analyze the received RTP packet stream, whereas they extract the data from the received RTCP receiver/sender reports (RR/SR) for the egress direction. RTCP application-defined (APP) packets transmit data that’s not directly available from these reports. In response to commands from the core, an agent will also adjusts the operational parameters of the VoIP implementations it supervises.

CQMC agent

VoIP implementation (SIP, H.323, and so on)

CQMC agent

VoIP implementation (SIP, H.323, and so on)

TCP/UDP/SCTP/IP network stack

Collected call data Processed call data quality metrics Operation control commands

The Core The core contains three components: data storage and retrieval, real-time data processing, and monitoring and control. The first component collects and caches the raw call data that’s necessary for the core to operate in both call directions. The most important data are the total number of received, lost, or late RTP packets, the RTP packet-stream jitter’s average and maximum values (from the start of the call), round-trip delay values, the playout buffer’s current size, the data for calculating the conditional loss probability (CLP), and a timestamp of the sampling instance. Using the collected raw call data, the second component calculates the metrics that the third component will use to evaluate call quality, including the current bit rates in both directions and the moving averages of the

IEEE INTERNET COMPUTING

Voice and call control/signaling

Figure 1. The Call Quality Monitoring and Control (CQMC) framework. The agents collect the call data and forward them to the core, which processes them and decides how to adapt the system’s operational parameters to improve each call’s voice quality (VQ). CQMC describes the necessary adaptation using a set of operation control commands that the core sends to the agents for execution. • • • •

CLP (burstiness), loss rates due to network or playout buffer loss, end-to-end delay, and network jitter.

Using a two-stage approach, the third component analyzes the calculated metrics and decides how to adapt the whole system’s operational parameters to improve each call’s quality. During the

www.computer.org/internet/

JULY • AUGUST 2005

37

The Road toward Internet Media

Delay update

Adjust interleaving Adjust frames/packet Adjust redundancy Adjust playout buffer

PLB update

Adjust frames/packet

Yes

Apply adjustment

Adjust redundancy Adjust frames/packet

Overall factor

No

deltaR > 0

Adjust interleaving

PLR update

Calculate expected deltaR

Reject adjustment

Return

Adjust playout buffer

Figure 2. Per-call metric analysis and operational parameter control flowchart. CQMC analyzes the packet-loss rate (PLR), delay (D), and loss burstiness (PLB) for each call. Upon voice quality (VQ) degradation, it selects one parameter adjustment as the system’s response. The system applies the selected adjustment only if its effect on VQ is positive (deltaR > 0). first stage, the third component maps the metrics of all active calls to call-quality figures based on the E-model’s R factor, which we calculate using its reduced form.9 Then, CQMC examines callquality metrics for all calls, looking for common trends in their variations (for example, quality degradation on the majority of the established calls). This operation is important for the second stage and controls whether the system implementing CQMC will transmit redundant information to protect each call’s media stream against loss. The system can use any kind of redundancy (such as FEC or redundant encoders) for loss recovery. Our approach to per-call metric analysis and operational parameter control — the second stage of call monitoring and control (see Figure 2) — is fundamentally different from other methods.4,5 First, we use all three IP network and system

38

JULY • AUGUST 2005

www.computer.org/internet/

impairments: packet-loss rate (PLR), end-to-end delay (D), and packet-loss burstiness (PLB). Furthermore, we can adjust four different parameters: redundancy, interleaving (that is, the spreading of adjacent voice frames to nonconsecutive packets to dispense burst-frame losses in a series of shorter, easier-to-conceal losses), the number of frames per packet, and playout buffer. Finally, rather than solving complex optimization problems, we determine solutions adaptively and with significantly lower complexity. As input variables, the adaptive algorithm uses the PLR, D, and PLB updates — which reflect changes in the encountered network conditions — to select an adjustment and decide whether to apply it. Based on a priori information3–5 and experimental data, we’ve imposed several constraints on the adaptation process to limit the adaptation space’s dimensions and increase the

IEEE INTERNET COMPUTING

Voice-Quality Monitoring

Implementation Platform e’ve implemented the Call Quality Monitoring and Control (CQMC) architecture as part of a complete IP private branch exchange (IP-PBX) system running on inAccess Networks’ MRG-110 residential gateway (www.inaccessnetworks.com/ ian/ian/products/reference_designs/MRG110-PB), which uses the Intel PXA255 processor and a Linux kernel. Support for analog telephony comes from an onboard digital signal processor as well as two foreign exchange station (FXS) and two foreign exchange office (FXO) ports. We implemented full-fledged telephony signaling (such as tone generation and detection, or caller ID), line echo cancellation (more than

W

50 decibels of Echo Return Loss Enhancement according to G.168/G.165), Automatic Gain Control (a 10-dB overall loudness rating [OLR] can be kept constant under varying public switched telephone network conditions), and standard bit-exact codecs (G.711, G.726, G.729A/B/AB) on Agere Systems’ DSP16410. The line-side devices (Infineon PEB3265/PEB4266 for the FXS and Silicon Laboratories SI3050/3019 for the FXO) fulfill all requirements on balance-return loss, noise levels, and quantization distortion (all the data is in 16-bit linear pulse-code modulation to minimize artifacts).These signal-quality metrics fully justify the use of the reduced form of R.1

algorithm’s speed and stability. For example, we deal with changes in burstiness only through adjustments to the number of frames per packet and interleaving (see Figure 2). During this stage, we first set a threshold on the running sum of PLR, D, or PLB that the agents periodically provide to the core. We determine this threshold with regard to the last update that forced an agent to apply an adjustment (as we explain later). We do this to discard trivial values, thus trading off the algorithm’s tracking ability with its long-term stability. Then, we select a single adjustment from the set that each update is allowed to affect, using a proprietary set of rules that involves a weighting strategy devised through extensive experimentation. Finally, we use smart indexing in precalculated lookup tables to decrease complexity in determining the selected adjustment’s effect on the PLR, D, and PLB. The CQMC core then independently maps the three deltas to expected changes of the R factor using piecewise linear approximations. These approximations’ summation yields the R factor’s total expected change, deltaR If deltaR is positive, indicating an expected improvement in call quality, the core applies the adjustment through the agents. Otherwise, the CQMC core ignores the adjustment. Due to the adaptive mechanism’s inherent feedback configuration, the applied adjustment will appear at its inputs within consequent updates. This can lead to oscillatory behavior. The core avoids such instabilities by suppressing

IEEE INTERNET COMPUTING

A modified version of the Asterisk softPBX (www.asterisk.org), which we’ve extended to support the hardware platform resources mentioned earlier, provides PBX functionality. Asterisk provides an H.323/SIP/MGCP-capable IP-PBX supporting a wide range of services and applications. The H.323 and RTP/RTCP stacks are based on the OpenH323 implementation (www.openh323.org), for which we developed a CQMC agent. Reference 1. R.G. Cole and J.H. Rosebluth,“Voice over IP Performance Monitoring,” ACM SIGCOMM Computer Comm. Rev., vol. 31, no. 2, 2001, pp. 9–24.

the deterministic effects of adjustments on future updates.

Measurements and Results We benchmarked CQMC using two MRG-110 platforms as the endpoints implementing the VQ monitoring and control mechanism (see the sidebar for a complete description of the implementation platform). The bottleneck in their interconnection was an asymmetric digital subscriber line (ADSL) connection with throughput 1 Mbps downstream and 256 Kbps upstream, which we deliberately selected to simulate a heavily loaded best-effort packet network. We used G.729A encoding, the most commonly used encoding in VoIP applications, on all voice streams during experimentation. To shed light on CQMC’s behavior in both heterogeneous and homogeneous network environments, we evaluated it using two scenarios. Concurrent Data and Voice Flows VoIP-packet flows demonstrate different behavior than data-packet flows. VoIP uses User Datagram Protocol (UDP) transport, so the applications must control the packet rate and its adaptation according to the loss and delay experienced throughout the network, whereas the TCP undertakes this task for data. To deal with packet loss variations within heterogeneous environments (concurrent TCP/UDP flows), our VoIP application can adjust both the number of voice frames in an RTP packet and the

www.computer.org/internet/

JULY • AUGUST 2005

39

80

1,000 800

60

600

40

400 20

200 0

1,000

2,000 3,000 4,000 Packet sequence number

0 5,000 6,000

200

80 150

40

50

20

0

0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 Packet sequence number R factor evolution

80 R factor

80 R factor

60

100

R factor evolution

60 40

60 40 20

20

(a)

100

100

100

0

Playout delay Total packet-loss rate (%)

Total packet-loss rate (%)

1,200

250 100

Playout delay Total packetloss rate (%)

Playout delay (milleseconds)

1,400

Total packet-loss rate (%)

Playout delay (milleseconds)

The Road toward Internet Media

1,000

2,000

3,000

4,000

Packet sequence number

0

5,000 6,000 (b)

1,000 2,000 3,000 4,000 5,000 6,000 7,000 Packet sequence number

Figure 3. System operation in a heterogeneous environment. (a) Without CQMC control, TCP flows produced by large data transfers lead to significant VQ degradation, mainly due to link congestion. (b) With CQMC control, rate adjustment and controlled transmission of redundancy for the voice flows, in conjunction with adapting the TCP data flows, greatly increase VQ. transmission of redundant voice frames under the CQMC agent’s control. To quantify CQMC’s performance, we tested it with two to three voice flows coexisting with two to three TCP dataflows produced by large FTP transfers. We first conducted the test without enabling the control logic; Figure 3a demonstrates the variation of the total packet loss, playout delay, and R of one voice flow. The packet loss was extremely high — averaging 16 percent throughout the call — and often higher than 20 percent. Because we use an adaptive jitter buffer, late losses were extremely low, so we don’t show them separately in the figure. At the packet with sequence number 2,800, several TCP flows were initiated, resulting in a tremendous deterioration of all metrics. The R factor dropped to zero, mainly due to the large delay. Soon afterward, the TCP flows adapted themselves, generating a small improvement, but the performance remains far from acceptable. We then conducted the same test with the control logic enabled (Figure 3b). Due to large packet

40

JULY • AUGUST 2005

www.computer.org/internet/

losses, CQMC increases the number of voice frames per RTP packet, decreasing the packet rate. Moreover, CQMC exploits the limited number of voice flows and activates redundancy (seq_num = 3,000), dramatically decreasing packet loss. The average packet loss throughout the call was 2 percent, which is reflected in R’s improvement. This behavior proves that in heterogeneous network environments, system performance relies on the reaction of the application controlling the voice flows and not solely on the adaptation of the TCP dataflows. Multiple Concurrent Voice Flows Operating CQMC is very different within a homogenous network environment that conveys only voice traffic; if the system applies redundancy to all voice flows, packet loss increases.11 To alleviate this detrimental effect, CQMC considers the number of voice flows and the level of packet loss of each one to decide whether it’s operating in a homogenous or heterogeneous environment. In the former case, CQMC avoids

IEEE INTERNET COMPUTING

2,500 2,000

100 80 60

1,500 40 1,000 20

500 0

Playout delay (milleseconds)

3,000

Playout delay Total packetloss rate (%)

Total packet-loss rate (%)

Playout delay (milleseconds)

3,500

0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 Packet sequence number

2,000 1,800 1,600 1,400 1,200 1,000 800 600 400 200 0

80 60 40 20

0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000 Packet sequence number R factor evolution

80

80

60

60

R factor

R factor

R factor evolution

40

40 20

20

0

100

100

100

(a)

Playout delay Total packetloss rate (%)

Total packet-loss rate (%)

Voice-Quality Monitoring

1,000 2,000 3,000 4,000 5,000 6,000 7,000 Packet sequence number

0 (b)

1,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000 Packet sequence sumber

Figure 4. System operation in a homogenous environment. (a) Without CQMC control, when the number of established calls exceeds the bottleneck link’s available bandwidth, the increased losses that each call experiences reduces quality to unacceptable levels. (b) With CQMC control, our mechanism detects whether the majority of the monitored calls’ VQ degrades due to common causes and uses — in this case, rate reduction as the main reaction to increased losses, resulting in rapid amelioration of VQ. redundancy, and the controlled parameter is the packet rate. To demonstrate this mechanism’s performance, we benchmark it in an environment with two to eight concurrent calls. We first conducted the test with the control logic disabled — Figure 4a presents the results for the duration of the first established call. After the fourth call’s initiation (seq_num = 2,200), the losses increase, and once the eighth call is in place (seq_num = 4,700), all the monitored call’s packets are lost and the delay is greater than 2.5 seconds for a 20-second time period. The R factor drops to zero, so the conversation can’t continue. We then conducted the same test with the control logic enabled (Figure 4b). Upon the fourth call’s initiation, the level of packet loss varies in much the same way it did with the control logic disabled. At this point, however, the system (under the control of CQMC) reacts, increasing the number of voice frames within each RTP packet for all active calls and reducing

IEEE INTERNET COMPUTING

the overall packet rate. The result is a gradual improvement of R.

O

ur system can’t, of course, overcome the need for better QoS management in the network backbone implementing IP telephony. To speed things up, however, ITSPs must be convinced to invest in this direction, and to do so, they must be able to create revenue using existing endpoint equipment to avoid further expenses. Our results indicate that an endpoint-based VQ monitoring and control strategy can provide significant earnings. The CQMC mechanism has several attributes that render it an eligible candidate for integration into the installed base of IP telephony systems today. It’s modular and can be easily partitioned among different processing elements; its core logic is independent of the VoIP protocol stack used; and finally, it presents low complexity and tracking capabilities because it is adaptive, performing equally well in different environments and under

www.computer.org/internet/

JULY • AUGUST 2005

41

The Road toward Internet Media

dynamically variable network conditions. Realtime VQ monitoring and control is an extremely active research area, and we plan to continue improving and extending CQMC, because we firmly believe that endpoint-based QoS is a mandatory interim step toward toll-quality VoIP. Acknowledgments We thank the inAccess Networks research department for providing the MRG-110 RGs used for the measurements. We also thank the anonymous reviewers.

References 1. D. Kagklis et al., “A Service and Network Management Framework for Providing Guaranteed QoS IP Services over WDM,” Proc. 6th IEEE High-Speed Network Management Circuit Conf. (HSNMC 03), LNCS 2720, Springer-Verlag, 2003, pp. 217–226. 2. C. Perkins, O. Hodson, and V. Hardman, “A Survey of Packet Loss Recovery Techniques for Streaming Audio,” IEEE Network, vol. 12, no. 5, 1998, pp. 40–48. 3. R. Ramjee et al., “Adaptive Playout Mechanisms for Packetized Audio Applications in Wide-Area Networks,” Proc. 13th Ann. Joint Conf. IEEE Computer and Comm. Societies (INFOCOM 94), IEEE CS Press, 1994, pp. 680–688. 4. C. Boutremans and J.-Y. Le Boudec, “Adaptive Joint Playout Buffer and FEC Adjustment for Internet Telephony,” Proc. 22nd Ann. Joint Conf. IEEE Computer and Comm. Societies (INFOCOM), IEEE CS Press, 2003, pp. 652–662. 5. C. Boutremans and J.-Y. Boudec, “Adaptive Delay-Aware Error Control for Internet Telephony,” Proc. 2nd IP Telephony Workshop, 2001, pp. 81–92. 6. Recommendation ITU-T P.800, Methods for Subjective Determination of Transmission Quality, Int’l Telecommunication Union, 1996. 7. Recommendation ITU-T P.862, Perceptual Evaluation of Speech Quality (PESQ), an Objective Method for End-toEnd Speech Quality Assessment of Narrowband Telephone Networks and Speech Codecs, Int’l Telecommunication Union, 2001. 8. Recommendation ITU-T G.107, The E-Model, a Computational Model for Use in Transmission Planning, Int’l Telecommunication Union, 2003. 9. J. Janssen et al., “Assessing Voice Quality in Packet-Based Telephony,” IEEE Internet Computing, vol. 6, no. 3, 2002, pp. 48–57. 10. R.G. Cole and J.H. Rosebluth, “Voice over IP Performance Monitoring,” ACM SIGCOMM Computer Comm. Rev., vol. 31, no. 2, 2001, pp. 9–24. 11. E. Altman, C. Barakat, and V. Ramos, “Queueing Analysis of Simple FEC Schemes for IP Telephony,” Proc. 20th Ann. Joint Conf. IEEE Computer and Comm. Societies (INFOCOM), IEEE CS Press, 2001, pp. 796–804.

42

JULY • AUGUST 2005

www.computer.org/internet/

Michael Manousos is a PhD student at the National Technical University of Athens (NTUA), a research associate at the Telecommunications Laboratory, and a senior software engineer for inAccess Networks. His research focuses on voice and data integration, IP telephony systems, and adaptive voice-quality control. Manousos has a diploma in electrical and computer engineering from NTUA. He is a member of the IEEE. Contact him at manousos@inaccess networks.com. Spyros Apostolacos is a PhD candidate at the National Technical University of Athens (NTUA) and a senior engineer at inAccess Networks. His research interests include efficient algorithms for high-capacity digital signal-processing systems for voice and fax processing. Apostolacos has a diploma in electrical and computer engineering from NTUA. Contact him at [email protected]. Ioannis Grammatikakis is a PhD candidate at the National Technical University of Athens (NTUA) and a senior engineer at inAccess Networks. His research interests include efficient digital signal-processing structures for voice processing. Grammatikakis has a diploma in electrical engineering and computer science from NTUA and an MSc in communications and signal processing from Imperial College. Contact him at [email protected]. Dimitrios Mexis is a senior engineer at inAccess Networks. His research interests include high-performance VoIP protocol stacks for embedded systems. Mexis has a diploma in electrical and computer engineering from the National Technical University of Athens and an MSc in communication systems and signal processing from the University of Bristol. Contact him at [email protected]. Dimitrios Kagklis is a PhD candidate at the National Technical University of Athens (NTUA) and a research associate at the Institute of Communications and Computer Systems (ICCS). His research interests include IP quality-of-service provisioning and service management. Kagklis has a diploma in electrical and computer engineering and an MSc in techno-economic systems from NTUA. He is a member of the IEEE and the ACM. Contact him at [email protected]. Efstathios Sykas is a professor in the Division of Communications, Electronics, and Information Engineering at the Department of Electrical Engineering at the National Technical University of Athens (NTUA). His research interests include third-generation personal and mobile communications. Sykas has a Dipl.-Ing. and a Dr.-Ing. in electrical engineering from NTUA. He is a member of the IEEE, the ACM, and the Technical Chamber of Greece. Contact him at [email protected].

IEEE INTERNET COMPUTING