Quality of Service, Resource and Energy Management

Introduction

Mobile computing systems operate under conditions that distinguish them from conventional wired networks: radio capacity varies with channel quality, users move between cells, devices have limited batteries, and applications compete for scarce spectrum, processing capacity, memory, backhaul bandwidth, and energy. A network may have high aggregate throughput yet provide poor user experience if packets arrive too late, jitter is excessive, handovers fail, or interactive traffic is delayed behind bulk transfers.

Quality of Service (QoS), resource management, and energy management are therefore interdependent design problems. QoS specifies the service performance required by applications and users. Resource management allocates radio, network, computing, and storage resources to satisfy those requirements. Energy management attempts to do so with minimum power expenditure while preserving service continuity, reliability, and security.

This chapter develops models and techniques for latency, throughput, reliability, availability, queueing, admission control, scheduling, radio-resource allocation, and battery-aware operation. It also examines the limitations of idealized models and the implications of virtualization, edge computing, artificial intelligence, and emerging 5G/6G systems.

Learning Objectives

After completing this chapter, the reader should be able to:

1. Distinguish QoS metrics from quality of experience (QoE) and relate both to application requirements.

2. Apply queueing, traffic, reliability, and availability models to mobile systems.

3. Explain radio-resource allocation, scheduling, admission control, and mobility-aware resource management.

4. Calculate link capacity, utilization, delay, battery lifetime, and availability under stated assumptions.

5. Evaluate trade-offs among throughput, latency, fairness, reliability, and energy consumption.

6. Design and critically assess a resource- and energy-management strategy for a mobile service.

1. QoS and QoE in Mobile Systems

QoS is the measurable performance delivered by a communication or computing service. Common QoS parameters include:

• Throughput: successfully delivered information per unit time, usually in bit/s.
• Goodput: useful application data delivered per unit time, excluding retransmissions and protocol overhead.
• Latency or delay: elapsed time between an event, such as packet generation, and its delivery or processing.
• Jitter: variation in packet delay, important for voice, video, and control applications.
• Packet-loss ratio: lost packets divided by transmitted packets over a defined observation interval.
• Reliability: probability that a service or transmission succeeds under specified conditions.
• Availability: proportion of time that a service is operational and accessible.
• Fairness: degree to which competing users receive an equitable service allocation.

QoE is the user's perceived quality. It depends on QoS but is not identical to it. For example, video quality may remain acceptable despite occasional packet loss when adaptive streaming is used; conversely, a low average delay may still produce poor QoE if delay spikes cause audio interruptions. QoE models may be application-specific, subjective, or objective. A useful engineering practice is to translate user-level requirements into measurable service-level objectives (SLOs), such as: “95% of interactive requests shall complete within 100 ms, and service availability shall exceed 99.9% per month.”

A service-level agreement (SLA) may include SLOs, measurement procedures, violation thresholds, remedies, and responsibilities. At postgraduate level, QoS engineering should be treated as a constrained optimization problem rather than as a search for maximum throughput alone.

2. Traffic and Service Models

2.1 Traffic characteristics

Mobile traffic is heterogeneous. Voice produces relatively small, periodic packets with strict delay constraints. Video generates high-rate streams whose rate may vary with encoding and scene complexity. Web traffic is bursty, while telemetry may consist of infrequent short messages. Traffic models must therefore identify:

• arrival rate, denoted by lambda, in packets/s or requests/s;
• packet-size distribution;
• burst duration and peak rate;
• correlation between arrivals;
• service-time distribution;
• tolerance to delay, loss, and reordering.

The average offered load is not sufficient for dimensioning. Two traffic sources with equal average rate can produce very different queue lengths if one is smooth and the other is highly bursty.

2.2 Link capacity and spectral efficiency

For a bandwidth-limited link with signal-to-interference-plus-noise ratio (SINR) gamma, an ideal upper bound on capacity is given by the Shannon expression

C = B log2(1 + gamma),

where C is capacity in bit/s, B is bandwidth in Hz, and gamma is a dimensionless linear SINR. This is an upper bound under ideal coding, arbitrarily long codewords, and suitable channel conditions. Practical systems achieve a lower rate because of finite block lengths, protocol overhead, imperfect channel estimation, interference, retransmissions, and implementation limitations.

If a system uses spectral efficiency eta in bit/s/Hz, practical throughput can be approximated as

R = B eta rho,

where rho is the fraction of time-frequency resources effectively available to the user. The factor rho accounts for scheduling, control overhead, reference signals, retransmissions, and sharing with other users.

SINR must not be inserted in decibels directly into the Shannon formula. If SINR is gamma_dB, its linear value is gamma = 10^(gamma_dB/10).

Worked Example 1: Capacity estimate

A user is assigned 10 MHz of bandwidth and experiences an SINR of 15 dB. The ideal capacity is estimated as follows:

1. Convert SINR: gamma = 10^(15/10) approximately 31.62.

2. Compute capacity: C = 10,000,000 log2(1 + 31.62).

3. Since log2(32.62) is approximately 5.03, C is approximately 50.3 Mbit/s.

This is an ideal upper bound. If scheduling and protocol efficiency together provide only 70%, an indicative useful rate is approximately 35.2 Mbit/s before considering retransmissions and variable channel conditions.

3. Queueing Models for Mobile Service Performance

Queueing models represent the interaction between arrivals, waiting buffers, and service facilities. A standard notation is Kendall's notation A/S/c, where A identifies the interarrival distribution, S the service-time distribution, and c the number of parallel servers. For example, M/M/1 denotes Poisson arrivals, exponentially distributed service times, and one server.

For an M/M/1 queue, let lambda be the mean arrival rate and mu the mean service rate, both in jobs/s. The utilization is

rho = lambda / mu.

The stability condition is rho < 1. If rho approaches one, mean delay increases sharply. Under the M/M/1 assumptions:

• mean number in the system: L = rho/(1-rho);
• mean time in the system: W = 1/(mu-lambda);
• mean number waiting: Lq = rho^2/(1-rho);
• mean waiting time in queue: Wq = rho/(mu-lambda).

These relationships follow from Little's law, L = lambda W, provided the system is stable and the averages are measured consistently. Little's law is more general than the M/M/1 model; it applies to many stable systems in steady state, including networks of queues, when L is the time-average population and lambda is the long-run throughput.

Worked Example 2: Queue delay

A mobile edge server receives 8 requests/s and processes 10 requests/s. Under an M/M/1 approximation, rho = 0.8. The mean time in the system is

W = 1/(10 - 8) = 0.5 s.

The mean waiting time excluding service is

Wq = 0.8/(10 - 8) = 0.4 s.

The mean number of requests in the system is L = lambda W = 8 x 0.5 = 4 requests. These are averages, not guarantees. A 500-ms mean delay does not imply that every request completes within 500 ms.

For multiple servers, an M/M/c model may be more appropriate. In a radio access network, however, service rates vary with channel quality and scheduling decisions, making simple stationary models approximate. Simulation, trace-driven analysis, or stochastic-network models may be needed for realistic evaluation.

4. Delay, Jitter, Loss, and Reliability

End-to-end latency can be decomposed into transmission, propagation, processing, queueing, retransmission, handover, and scheduling components:

D_total = D_tx + D_prop + D_proc + D_queue + D_retx + D_handover.

The decomposition is conceptual; some terms may overlap depending on the measurement boundary. Transmission delay for a packet of L bits over a rate R bit/s is L/R seconds. Propagation delay is distance divided by propagation speed, although in mobile networks it is often much smaller than queueing and processing delays. Retransmissions improve reliability but increase delay and energy consumption.

If independent transmission attempts each succeed with probability p and at most k attempts are allowed, the probability of at least one successful attempt is

P_success = 1 - (1-p)^k.

This formula assumes independent attempts and constant success probability. In a fading channel, attempts may be correlated, so the formula can be optimistic.

For a path with independent component availabilities A1, A2, ..., An in series, end-to-end availability is

A_series = product from i=1 to n of Ai.

For redundant parallel components that operate independently, availability is

A_parallel = 1 - product from i=1 to n of (1-Ai).

Independence is a strong assumption. Common power, backhaul, software, or environmental failures can invalidate these simple expressions.

Availability is also linked to mean time between failures (MTBF) and mean time to repair (MTTR). Under a basic alternating renewal approximation,

A = MTBF/(MTBF + MTTR).

This relation assumes well-defined operating and repair cycles and is not a substitute for a complete failure model.

Worked Example 3: Availability

A mobile edge service has MTBF of 1,000 hours and MTTR of 2 hours. Its approximate availability is

A = 1000/(1000+2) = 0.998004,

or about 99.80%. If a contractual target is 99.9%, improving only mean repair time may be necessary but not sufficient; redundancy, fault detection, and reduction of correlated failures should also be considered.

5. Resource Management Architecture

Mobile resource management operates across several layers:

1. Radio layer: spectrum, time-frequency resource blocks, transmit power, modulation, coding, beam selection, and hybrid automatic repeat request.

2. Transport and network layer: routing, congestion control, slicing, backhaul capacity, and mobility anchoring.

3. Edge and cloud layer: CPU cycles, memory, accelerators, storage, and service placement.

4. Device layer: processor frequency, radio state, sensors, display, and application execution.

5. Policy layer: priorities, SLAs, charging, fairness, security, and regulatory restrictions.

A resource manager may optimize an objective such as weighted utility:

maximize sum over users i of wi Ui(ri)

subject to capacity, power, latency, reliability, and fairness constraints. Here ri is the allocated rate, wi is a policy weight, and Ui is a utility function. A logarithmic utility can promote proportional fairness, while a linear utility tends to favor aggregate throughput. The choice of utility is a policy decision, not merely a mathematical detail.

6. Scheduling and Admission Control

6.1 Scheduling policies

Common scheduling policies include:

• First-come, first-served: simple and transparent, but insensitive to urgency and channel quality.
• Round robin: provides time sharing but may waste resources on users with poor instantaneous efficiency.
• Maximum-throughput scheduling: favors users with favorable channels; efficient but potentially unfair.
• Proportional fair scheduling: balances instantaneous rate and historical service, often represented conceptually by selecting users with a high ratio of achievable rate to average past rate.
• Priority or deadline scheduling: useful for emergency, industrial, or control traffic, but can starve lower-priority users.
• Earliest-deadline-first: serves the packet with the nearest deadline; effective when deadlines are known and schedulable.
• QoS-aware hybrid scheduling: combines channel quality, queue length, delay, reliability, and service class.

No scheduler dominates in every environment. A throughput-maximizing scheduler may reduce fairness; a strict latency policy may reduce spectral efficiency; and a power-aware policy may sacrifice instantaneous rate to avoid energy-intensive retransmissions.

6.2 Admission control

Admission control determines whether a new flow or session can be accepted without violating existing guarantees. A simple deterministic rule accepts a flow only if the predicted demand plus current reservation does not exceed capacity. A statistical rule may accept traffic based on a target probability of congestion or delay violation.

Admission control is essential for hard QoS services. Without it, each flow may appear acceptable in isolation while aggregate demand causes unacceptable delay for all users. Parameters may include peak rate, mean rate, burst size, latency bound, packet-loss target, mobility risk, and required redundancy.

6.3 Mobility-aware management

Handover prediction can reserve resources in a target cell before a user arrives. This reduces handover interruption but may leave reserved capacity unused. A robust strategy estimates prediction confidence, user velocity, route, target-cell load, and service criticality. Admission decisions should account for both current cell resources and likely future cells.

7. Energy Consumption and Battery Management

A mobile device consumes energy in its radio interface, processor, display, memory, sensors, and peripheral components. Radio energy is not simply proportional to transmitted bits. State transitions, synchronization, idle listening, retransmissions, and signaling may dominate for small or sporadic transfers.

A first-order energy model is

E_total = E_tx + E_rx + E_cpu + E_display + E_idle + E_overhead.

For a component operating at approximately constant power P for duration t, E = Pt. If power varies, energy is the integral of power over time:

E = integral from 0 to T of P(t) dt.

Battery life can be estimated by

T_life = E_batt / P_avg,

where E_batt is usable battery energy in joules and P_avg is average device power in watts. This is an engineering approximation. Real batteries exhibit rate dependence, temperature effects, aging, voltage variation, and a difference between nominal and usable capacity.

Energy per bit is a useful communication metric:

E_bit = E_comm / N,

where E_comm is communication energy and N is the number of successfully delivered useful bits. It should be distinguished from energy per transmitted bit, which may include retransmissions and protocol overhead.

Energy-saving mechanisms

• Discontinuous reception (DRX): allows the radio to sleep during predictable inactivity intervals.
• Batching: combines small transfers to reduce repeated state transitions, at the cost of delay.
• Adaptive transmission power: reduces interference and energy when link conditions permit.
• Rate and modulation adaptation: uses robust low-rate modes in poor channels and efficient modes in good channels.
• Computation offloading: shifts processing to edge or cloud resources, but transmission energy and privacy risks must be included.
• Caching: avoids repeated downloads and reduces radio use.
• Location and sensing adaptation: changes sampling frequency according to application need.
• Wake-up radios and low-power modes: support event-driven operation with reduced idle consumption.

The optimal policy depends on the objective. Minimizing device energy may increase network energy or latency. Consequently, energy management should consider system-wide energy and carbon impact rather than only handset battery drain.

Worked Example 4: Battery lifetime

A device has a usable battery energy of 36,000 J and average power consumption of 1.5 W. The idealized lifetime is

T_life = 36,000/1.5 = 24,000 s,

which is 6.67 hours. If an energy-saving policy reduces average power to 1.0 W, the estimated lifetime becomes 10 hours. The improvement is significant, but the estimate assumes constant average power and does not model battery aging or thermal limits.

8. Cross-Layer Optimization and Edge Computing

Cross-layer management allows information to move between protocol layers. For example, an application can request a latency budget, the transport layer can expose congestion state, the radio scheduler can report channel conditions, and the edge platform can migrate a service closer to the user. This can improve efficiency but increases architectural coupling, policy complexity, and security exposure.

Edge computing reduces propagation and cloud-processing delay by placing computation near users. Nevertheless, total latency includes radio access, transport, queueing, service execution, and return-path components. Offloading is beneficial only when the communication and coordination overhead is less than the local-computation cost and when the edge service has sufficient capacity.

A practical offloading decision should compare:

• local execution time and energy;
• uplink and downlink transmission time;
• remote execution time;
• radio and device energy;
• privacy and trust requirements;
• availability of the edge node;
• migration and state-synchronization cost.

9. AI-Assisted Resource and Energy Management

Machine learning can predict traffic demand, channel quality, handover probability, battery state, and application workload. These predictions may support proactive slicing, cache placement, scheduling, and sleep control. Reinforcement learning can learn policies for sequential decisions such as selecting a serving cell or transmission mode.

However, AI does not remove the need for engineering constraints. A learned scheduler may violate latency or fairness requirements under distribution shift. Training data can contain bias, models may be vulnerable to manipulation, and decisions may be difficult to explain. Safe deployment should use constrained optimization, fallback policies, monitoring, uncertainty estimation, and conservative exploration. Energy used to train and run the model must also be included in sustainability assessments.

10. Illustrative Case Study: Low-Latency Mobile Health Monitoring

Background

Consider an illustrative mobile health-monitoring service in which wearable sensors transmit heart-rate and motion data to a smartphone and then to an edge server. Routine data can tolerate several seconds of delay, but an alert associated with a possible cardiac event should reach a clinician-facing service within 200 ms with high reliability.

Problem

The system experiences three difficulties: routine sensor uploads compete with alerts, smartphone radio sleep cycles introduce delay, and users move between cells. Simply increasing average throughput does not guarantee alert performance.

Analysis

The service is divided into two classes. Alert traffic receives a strict latency and reliability target; routine traffic uses elastic delivery. The radio and edge queues are monitored separately because a low radio delay can be followed by edge-server congestion. Handover probability is estimated from device location and velocity. Battery measurements show that frequent keep-alive messages and radio state transitions are significant energy consumers.

Approach

The proposed design uses:

1. priority admission for alert traffic with reserved capacity;

2. deadline-aware scheduling for alert packets;

3. packet duplication over carefully selected independent paths when the reliability requirement justifies the energy cost;

4. batching of routine sensor data;

5. adaptive DRX with an immediate wake-up mechanism for alerts;

6. edge placement near the serving access network;

7. predictive handover preparation for high-velocity users;

8. continuous monitoring of 95th- and 99th-percentile delay rather than only the mean.

Suppose a single alert transmission attempt succeeds with probability 0.98 and two attempts are sufficiently independent. Then the success probability is

1 - (1-0.98)^2 = 1 - 0.0004 = 0.9996,

or 99.96%. The independence assumption must be tested; if both attempts share the same congested backhaul, duplication will not provide this improvement.

Outcome

The design improves alert reliability and tail latency while reducing routine-data energy through batching. It does, however, consume additional energy during alert duplication and requires careful privacy protection because health data is sensitive.

Learning points

The case demonstrates that QoS classes, admission control, scheduling, mobility prediction, and energy policies must be jointly designed. It also shows why mean performance is inadequate for critical applications: tail delay, correlated failures, and battery impact can determine actual service quality.

11. Challenges, Governance, and Best Practices

Technical limitations

Analytical models often assume Poisson arrivals, independent failures, stationary service rates, or perfect channel knowledge. Mobile systems violate these assumptions through correlated traffic, fading, mobility, interference, software updates, and shared infrastructure. Models should therefore be validated using measurements, trace-driven simulation, field trials, and sensitivity analysis.

Security and privacy

Resource-management data may reveal location, health status, application use, or enterprise activity. Admission and prioritization policies can also create denial-of-service opportunities. Security controls should include authentication, authorization, isolation between slices, encrypted telemetry, rate limiting, and auditability. Privacy-preserving measurement should collect only the information necessary for optimization.

Fairness and governance

Priority mechanisms must be transparent and justified. Emergency traffic may appropriately pre-empt ordinary traffic, but commercial prioritization can create discriminatory access. Regulators and service providers should define neutrality, accessibility, lawful interception, emergency-service treatment, and data-retention obligations. Fairness should be evaluated across users, locations, device capabilities, and application classes.

Recommended engineering practice

1. Define application-level SLOs before selecting network metrics.

2. Measure distributions and tail percentiles, not only averages.

3. Separate capacity planning from real-time scheduling.

4. Validate assumptions about independence, stationarity, and channel knowledge.

5. Include retransmissions, signaling, sleep transitions, and control overhead in energy accounting.

6. Use admission control for services with hard guarantees.

7. Design for mobility and correlated failures.

8. Test under congestion, poor coverage, handover, and battery-aging conditions.

9. Use explainable and constrained AI when automated control is introduced.

10. Report trade-offs among throughput, delay, reliability, fairness, cost, and energy.

Contemporary Developments

5G and emerging 6G systems extend QoS management through network slicing, massive antenna systems, millimeter-wave and terahertz research, integrated sensing and communications, ultra-reliable low-latency communication, and cloud-native core networks. Multi-access edge computing supports computation placement close to users, while non-terrestrial networks provide broader coverage but introduce propagation and mobility challenges. Digital twins and closed-loop automation may enable predictive management, although their accuracy and operational safety remain research concerns.

Future systems are likely to optimize communication, computation, sensing, and energy jointly. Open problems include reliable performance under AI control, energy-aware semantic communication, carbon-aware orchestration, federated learning with strict latency requirements, and resource allocation across terrestrial, aerial, and satellite networks.

Chapter Summary

Mobile QoS is multidimensional: throughput, delay, jitter, loss, reliability, availability, fairness, and energy all influence service quality. Queueing theory explains why utilization near full capacity can produce disproportionate delay, while Little's law links population, throughput, and time. Capacity estimates based on Shannon's expression are useful bounds but must be reduced for practical overhead and imperfect channels. Reliability and availability calculations require explicit assumptions about independence and failure behavior.

Resource management spans radio, transport, edge, cloud, device, and policy layers. Scheduling policies trade throughput against fairness, urgency, and energy, while admission control protects existing guarantees from overload. Energy management uses sleep modes, batching, adaptive transmission, caching, offloading, and workload control, but every saving may impose a latency, reliability, or network-energy cost. AI can improve prediction and control only when constrained, monitored, and evaluated for safety, privacy, and sustainability. Effective postgraduate-level design therefore treats QoS, resource allocation, mobility, reliability, and energy as a coupled systems problem.

Key Terms

QoS: Measurable performance characteristics of a service.

QoE: User-perceived quality of an application or service.

Throughput: Successfully delivered data per unit time.

Goodput: Useful application data rate excluding retransmissions and overhead.

Latency: Time between an event and its completion or delivery.

Jitter: Variation in packet delay.

Reliability: Probability that a specified operation succeeds.

Availability: Fraction of time a service is operational and accessible.

SINR: Ratio of desired signal power to interference plus noise power.

Spectral efficiency: Data rate delivered per unit bandwidth.

Queue utilization: Ratio of arrival rate to service rate in a queue.

Little's law: Relationship L = lambda W between average population, throughput, and time.

Admission control: Decision process for accepting or rejecting new flows.

Scheduling: Allocation of service opportunities among competing users or packets.

Proportional fairness: Allocation principle balancing efficiency and equitable service.

DRX: Discontinuous reception that allows a radio to sleep periodically.

Energy per bit: Communication energy divided by useful delivered bits.

Edge computing: Processing performed near the data source or user.

Network slicing: Logical partitioning of shared infrastructure for different service requirements.

Service-level objective: Quantified performance target for a service.

Review Questions

Short Answer Questions

1. Distinguish QoS from QoE and give one example of their interaction.

2. State the stability condition for an M/M/1 queue and explain its significance.

3. Why must SINR in decibels be converted before using the Shannon capacity equation?

4. Define availability and identify the roles of MTBF and MTTR.

5. List four techniques used to reduce mobile-device energy consumption.

Long Answer Questions

1. Explain how throughput, latency, jitter, reliability, and fairness interact in mobile-network design.

2. Derive and interpret the principal M/M/1 queue relationships, including their assumptions and limitations.

3. Compare round-robin, maximum-throughput, proportional-fair, and deadline-aware scheduling.

4. Discuss cross-layer resource management and its benefits, risks, and implementation challenges.

5. Explain why energy management should be evaluated at device, network, and system levels.

Analytical / Application Questions

1. A service receives 12 requests/s and has a mean service rate of 15 requests/s. Using an M/M/1 model, calculate utilization, mean time in the system, and mean number of requests in the system. Interpret the result.

2. Design an admission-control and scheduling policy for a mobile industrial-control service that requires low latency and high reliability while sharing resources with video traffic.

3. Evaluate whether computation offloading is beneficial for a battery-powered device. Identify the measurements and assumptions required, and discuss privacy, latency, and energy trade-offs.

Multiple Choice Questions with Answers

Question 1

For an M/M/1 queue, what condition is required for stability?

A. lambda is greater than mu

B. lambda equals zero

C. lambda is less than mu

D. rho is greater than one

Correct Answer: C. lambda is less than mu

Question 2

A bandwidth of 10 MHz and an SINR of 15 dB give an ideal Shannon capacity closest to:

A. 5.03 Mbit/s

B. 15.0 Mbit/s

C. 35.2 Mbit/s

D. 50.3 Mbit/s

Correct Answer: D. 50.3 Mbit/s

Question 3

Which metric represents useful application data delivered per unit time, excluding retransmissions?

A. Goodput

B. Jitter

C. Availability

D. Propagation delay

Correct Answer: A. Goodput

Question 4

If an M/M/1 queue has arrival rate 8 requests/s and service rate 10 requests/s, its mean time in the system is:

A. 0.05 s

B. 0.2 s

C. 0.5 s

D. 5 s

Correct Answer: C. 0.5 s

Question 5

Which scheduling policy explicitly balances instantaneous achievable rate with historical service?

A. First-come, first-served

B. Proportional fair

C. Strict round robin

D. Random scheduling

Correct Answer: B. Proportional fair

Question 6

With MTBF of 1000 hours and MTTR of 2 hours, approximate availability is:

A. 90.0%

B. 98.0%

C. 99.80%

D. 99.99%

Correct Answer: C. 99.80%

Question 7

Which technique primarily reduces energy by allowing the radio to sleep during predictable inactivity?

A. Discontinuous reception

B. Packet fragmentation

C. Maximum-throughput scheduling

D. Increasing retransmissions

Correct Answer: A. Discontinuous reception

Question 8

For two independent transmission attempts each having success probability 0.98, the probability of at least one success is:

A. 0.9604

B. 0.9800

C. 0.9980

D. 0.9996

Correct Answer: D. 0.9996

Question 9

What is the principal purpose of admission control?

A. To maximize packet size in every flow

B. To prevent new traffic from violating existing guarantees

C. To eliminate all handovers

D. To replace encryption mechanisms

Correct Answer: B. To prevent new traffic from violating existing guarantees

Question 10

Which statement about the Shannon capacity expression is correct?

A. It is an exact practical throughput for every radio system

B. It uses SINR in decibels without conversion

C. It is an ideal upper bound under stated assumptions

D. It applies only to optical networks

Correct Answer: C. It is an ideal upper bound under stated assumptions