QoS and Resource Management

Handling multimedia streamsSearch on YouTubeBookmarkSaved

There are two major approaches to handle multimedia streams in computing and communication environments.

Need: QoS and resource management concepts are crucial to meet the heterogeneity of requirements that different distributed multimedia applications bring.

Requirements and constraintsSearch on YouTubeBookmarkSaved

(Text book T1, chapter 2.1.) To be accepted, multimedia systems must satisfy very heterogeneous requirements under different conditions and constraints of their run-time computing and communication environment. The central requirement is real time with deadlines.

Real timeSearch on YouTubeBookmarkSaved

The German National Institute for Standardization (DIN, similar to ANSI) defines it:

A real-time process is a process which delivers the results of the processing in a given time-span.

It further says:

• Real-time systems must meet strict timing deadlines to operate correctly.
• The system must handle both deterministic and stochastic events, and enforce both internal and external time constraints.

DeadlinesSearch on YouTubeBookmarkSaved

A deadline is the latest acceptable time limit for presenting a processing or communication result. It marks the border between normal (correct) and anomalous (failing) behaviour.

Characteristics of real-time systemsSearch on YouTubeBookmarkSaved

Real-time requirements on multimedia systemsSearch on YouTubeBookmarkSaved

Each unit of audio or video must be presented by a well-determined deadline. To meet the timing of continuous media, scheduling techniques are used for every resource on the path: CPU, main memory, storage, I/O devices, networks.

• Fault tolerance: less strict than traditional real-time systems. A short failure has little impact. Exceptions exist, such as remote surgery.
• Missing a deadline: an occasional miss is barely visible in video (one lost frame). Audio is different: human ears are more sensitive to small glitches.
• Periodic sampling: all continuous media signals are sampled periodically. This timing must be synchronized through the whole system.
• Bandwidth demand: not stringent. It is negotiable, because most streaming services use ABR.
• Multimedia systems must also meet the requirements of external devices, for example an orchestra.

ConstraintsSearch on YouTubeBookmarkSaved

• Data movement constraint
• Segmentation and reassembly
• Error-recovery mechanisms

Concepts of QoSSearch on YouTubeBookmarkSaved

(Text book T1, chapter 2.2.) Quality of Service is the defined and controlled behaviour of a service, expressed through quantitative, measurable parameters. The ISO standard defines QoS as a concept for specifying how "good" the offered networking services are.

The key: because requirements are heterogeneous, services in multimedia systems must be parameterized.

QoS layeringSearch on YouTubeBookmarkSaved

Traditional QoS was defined for the network layer and later extended to the transport layer. End-to-end quality also depends on other areas, so QoS is defined in a layered architecture.

Service objectsSearch on YouTubeBookmarkSaved

Service objects are the entities on which services are performed.

• Media sources: the original source of the content (a video file).
• Connections: communication channels between devices (a TCP connection).
• Virtual circuits (VCs): logical paths through a network (a circuit-switched connection).

QoS parameterizationSearch on YouTubeBookmarkSaved

QoS parameterization defines specific quality requirements for each service object, so the service meets the expected standards. Different bodies and systems parameterize it differently:

QoS specificationSearch on YouTubeBookmarkSaved

Perceptual QoSSearch on YouTubeBookmarkSaved

Two major requirements:

• Description of perceptive qualities: media quality, window size, response time, security.
• Description of pricing qualities: the range of price the user is willing to pay.

Application QoSSearch on YouTubeBookmarkSaved

Requirements are stated in terms of:

• Media quality: frame rate, frame resolution, latency, jitter.
• Media relations: media transformations, frame syncs.
• Adaptation rules: actions to take if network bandwidth is scarce.

Application-level specification approachesSearch on YouTubeBookmarkSaved

System QoS parametersSearch on YouTubeBookmarkSaved

These describe requirements placed on communication and computing services. They are derived from application QoS.

• Quantitative: bits per second, number of errors, task processing time, task period.
• Qualitative: frame synchronization, ordered data delivery, error-recovery mechanisms, scheduling mechanisms.

Network QoS parametersSearch on YouTubeBookmarkSaved

These describe requirements on low-level network services.

• Network load: inter-arrival time on the connection, burstiness, packet or cell size, service time in the node for a connection's packet.
• Network performance: delay, packet loss, jitter.

All these parameters are enclosed in the traffic envelope.

Device QoS parametersSearch on YouTubeBookmarkSaved

These specify timing and throughput demands for media data units from audio and video devices, such as speakers or video capture boards.

QoS parameter values and service classesSearch on YouTubeBookmarkSaved

The QoS parameter values decide the service class: guaranteed, predictive (predictable) or best effort. The class determines two things: the reliability of the offered QoS and the utilization of resources.

Class 1: guaranteed servicesSearch on YouTubeBookmarkSaved

Guaranteed services make sure multimedia applications receive the resources they need (bandwidth, processing power) to work without interruption. For multimedia this usually means the right bandwidth, low delay and minimal errors. This is resource reservation.

Deterministic representationCopy Anchor Link

Deterministic QoS means performance parameters can be calculated or guaranteed in advance. They are positive real numbers (frame rate, resolution) measured over time $T$:

$$QoS : T \rightarrow R$$

Here $T$ is the time duration of the service for which QoS is specified, and $R$ is the set of positive real numbers representing the QoS parameters.

Sometimes one value cannot describe the requirement ("How many dosa do you eat?"). So bounds come as a single value, a pair or a triple.

The pair specification splits the range into regions:

• Acceptable quality region: $[QoS_{min}, QoS_{max}]$
• Unacceptable quality region: $QoS(t) < QoS_{min}$

The deck's graph plots frame rate (fps) against resolution (pixels). Dashed lines mark the minimum and maximum QoS frame rate (10 and 30 fps) and the minimum and maximum resolution (80 x 40 and 640 x 480). The area inside is acceptable. A very low frame rate or resolution falls in the unacceptable region: choppy or pixelated video, like 240p for a viewer who expects HD.

The summary also names a statistical representation for guaranteed services, next to the deterministic one.

Class 2: predictable servicesSearch on YouTubeBookmarkSaved

Predictable service gives a consistent level of performance based on historical data and predictive algorithms.

1. Historical data collection: gather bandwidth, latency, jitter and packet loss over a period.
2. Data analysis: find patterns, trends and anomalies with mean, median, standard deviation and correlation.
3. Predictive model development: a simple statistical model or a complex machine learning algorithm.
4. Prediction: use current conditions and history to predict future performance.
5. QoS guarantee: promise a level based on the prediction, such as minimum bandwidth, maximum latency, or both.

Class 3: best-effort servicesSearch on YouTubeBookmarkSaved

Best-effort services guarantee no level of performance or reliability. There are no promises about speed, latency or delivery. The network tries its best.

Multimedia services (video streaming, online gaming) need stable, predictable QoS. A buffering stream or a lagging game hurts the user badly. So they need guaranteed bandwidth, delay and jitter, not best effort.

Resource managementSearch on YouTubeBookmarkSaved

(Text book T1, chapter 2.3.) Available resources decide the actual QoS. A resource is a system entity that tasks need to manipulate data. Each resource has distinguishing characteristics.

Classification of resourcesSearch on YouTubeBookmarkSaved

Capacity and requirementSearch on YouTubeBookmarkSaved

• Capacity: each resource needs a certain capacity to complete a task: processor bandwidth, network bandwidth or memory space.
• Requirement: real-time multimedia needs resources not only in large amounts but exactly when required.

What resource management meansSearch on YouTubeBookmarkSaved

In multimedia systems, resource management makes sure the system has enough CPU, memory, storage and network bandwidth for the audio, video and data being processed.

Why it matters: audio and video files are large and need fast processing. Compression helps, but high-quality audio and video together still need many resources. A movie must load quickly with sound matching the video, without delay or buffering. Poor management gives delays, pixelation or late parts.

PhasesSearch on YouTubeBookmarkSaved

Requirements on resource managementSearch on YouTubeBookmarkSaved

Media server architectureSearch on YouTubeBookmarkSaved

(Text book T1, chapter 4.1.) A media server is a special file server that sends multimedia data (video, music, time-sensitive content) to users within a specific time frame. It must deliver smoothly for a good experience, such as a movie stream or live broadcast. Media servers vary in hardware and software.

Normal file server versus multimedia file serverSearch on YouTubeBookmarkSaved

Pull and push modelsSearch on YouTubeBookmarkSaved

Quick self-testSearch on YouTubeBookmarkSaved

Compare scaling/adaptation with resource reservation. Give one example of each.

Scaling adapts media quality in real time to the resources available (ABR on Netflix, YouTube). Reservation reserves bandwidth, CPU and memory in advance so they are always available (VoIP: Skype, WhatsApp, Zoom).

Define a deadline. Differentiate soft and hard deadlines.

A deadline is the latest acceptable time to present a result; it separates correct from failing behaviour. Soft deadlines can be missed without catastrophe if not too many are missed and not by much. Hard deadlines must never be violated; a violation is a system failure, and they come from the physical characteristics of the process.

Why is missing a deadline worse for audio than for video?

An occasional missing video frame is barely visible. Human ears are more sensitive to small sound glitches, so a missed audio deadline is noticed.

List the five QoS layers with two parameters each.

Perceptual (visual quality, smoothness), Application (latency, jitter), System (processing power, memory), Device (display resolution, battery life), Network (bandwidth, packet loss).

A mobile video-learning app must specify its QoS. Which application-level specification approach would you use and which parameters?

A parameter-based approach (QoS in data structures) with adaptation rules, or a control-based approach for adaptive policies. Parameters: media quality (frame rate, resolution, latency, jitter), media relations (audio-video sync), and adaptation rules (lower bitrate when bandwidth is scarce).

Explain the pair-of-values specification and its regions.

It gives a minimum and maximum: $QoS_{min} \le QoS(t) \le QoS_{max}$, for example 15 to 30 fps. Values inside form the acceptable region; $QoS(t) < QoS_{min}$ is the unacceptable region (choppy or pixelated video).

Bandwidth samples are 10, 12, 8, 11, 9 Mbit/s. Find the predictable bandwidth with L = 3.

Average the last three: $(8 + 11 + 9)/3 = 9.33$ Mbit/s, about 9 Mbit/s as on the slide.

Why is best effort not enough for multimedia?

Best effort gives no guarantee on speed, latency or delivery. Streaming and gaming need stable bandwidth, delay and jitter; otherwise buffering or lag ruins the experience.

Classify resources with examples.

Active (CPU, GPU, network adapter) versus passive (RAM, disk). Exclusive (camera while recording) versus shared (common storage). Single (one GPU) versus multiple (multi-core CPU, several GPUs).

Name the two phases and four requirements of resource management.

Phases: establishment (admission check before starting) and runtime (continuous allocation and fair adjustment). Requirements: throughput, delay (local and end-to-end), jitter, reliability.