• Aucun résultat trouvé

Variable Bit Rate (VBR)

Dans le document Ivan Marsic (Page 193-197)

Some signal-compression techniques convert a signal into a bit stream that has variable bit rate (VBR). For instance, MPEG-2 is a family of standards for such variable bit rate compression of video signals. The bit rate is larger when the scenes of the compressed movies are fast moving than when they are slow moving. Direct Broadcast Satellite (DBS) uses MPEG-2 with an average rate of 4 Mbps.

To specify the characteristics of a VBR stream, the network engineer specifies the average bit rate and a statistical description of the fluctuations of that bit rate. More about such descriptions will be said later.

Messages

Many user applications are implemented as processes that exchange messages over a network. An example is Web browsing, where the user sends requests to a web server for Web pages with embedded multimedia information and the server replies with the requested items. The message traffic can have a wide range of characteristics. Some applications, such as email, generate isolated messages. Other applications, such as distributed computation, generate long streams of messages. The rate of messages can vary greatly across applications and devices.

To describe the traffic characteristics of a message-generating application, the network engineer may specify the average traffic rate and a statistical description of the fluctuations of that rate, in a way similar to the case of a VBR specification.

See definition of fidelity in:

B. Noble, “System support for mobile, adaptive applications,” IEEE Personal Communications, 7(1), pp.44-49, February 2000.

E. de Lara, R. Kumar, D. S. Wallach, and W. Zwaenepoel, “Collaboration and Multimedia Authoring on Mobile Devices,” Proc. First Int’l Conf. Mobile Systems, Applications, and Services (MobiSys 2003), San Francisco, CA, pp. 287-301, May 2003.

In any scenario where information is communicated, two key aspects of information are fidelity and timeliness. Higher fidelity implies greater quantity of information, thus requiring more resources. The system resources may be constrained, so it may not be possible to transmit, store, and visualize at a particular fidelity. If memory and display are seen only as steps on information’s way to a human consumer, then they are part of the communication channel. The user could experience pieces of information at high fidelity, sequentially, one at a time, but this requires time and, moreover, it requires the user to assemble in his or her mind the pieces of the puzzle to experience the whole. Some information must be experienced within particular temporal and or spatial (structural?) constraints to be meaningful. For example, it is probably impossible to experience music one note at a time with considerable gaps in between. Or, a picture cannot be experienced one pixel at a time. Therefore, the user has to trade fidelity for temporal or spatial capacity of the communication channel.

Time (sec) 0.007 0.4

0

0.4

Analog speech waveform Sampled signal Sampled & quantized signal

Amplitude

Figure 3-1: Analog speech signal sampling, and quantization to 4 bits.

Information loss may sometimes be tolerable; e.g., if messages contain voice or video data, most of the time the receiver can tolerate some level of loss.

Shannon had to introduce fidelity in order to make problem tractable [Shannon & Weaver 1949].

Information can be characterized by fidelity ~ info content (entropy). The effect of a channel can be characterized as deteriorating information’s fidelity and increasing the latency:

fidelityIN + latencyIN→()_____)→ fidelityOUT + latencyOUT

Wireless channels in particular suffer from limitations reviewed in Volume 2. Increasing the channel capacity to reduce latency is usually not feasible—either it is not physically possible or it is too costly.

Information qualities can be considered in many dimensions. We group them in two opposing ones:

• Those that tend to increase the information content

• Delay and its statistical characteristics

The computing system has its limitations as well. If we assume finite buffer length, then in addition to delay problem, there is a random loss problem. This further affects the fidelity.

Fidelity has different aspects, such as:

• Spatial (sampling frequency in space and quantization – see Brown&Ballard CV book)

• Temporal (sampling frequency in time)

• Structural (topologic, geometric, …)

Delay or latency may also be characterized with more parameters than just instantaneous value, such as the amount of variability of delay, also called delay jitter. In real life both fidelity and latency matter and there are thresholds for each, below which information becomes useless. The system is forced to manipulate the fidelity in order to meet the latency constraints. A key question is, how faithful should signal be in order to be quite satisfactory without being too costly? In order arrive at a right tradeoff between the two, the system must know:

1. Current channel quality parameters, e.g., capacity, which affect fidelity and latency 2. User’s tolerances for fidelity and latency

The former determines what can be done, i.e., what fidelity/latency can be achieved with the channel at hand, and the latter determines how to do it, i.e., what matters more or less to the user at hand. Of course, both channel quality and user preferences change with time.

Example with telephone: sound quality is reduced to meet the delay constraints, as well as reduce the costs.

Targeted reduction of information fidelity in a controlled manner helps meet the latency constraints and averts random loss of information. Common techniques for reducing information fidelity include:

• Lossless and lossy data compression

• Packet dropping (e.g., RED congestion-avoidance mechanism in TCP/IP)

• …?

The above presentation is a simplification in order to introduce the problem. Note that there are many other relevant parameters, such as security, etc., that characterize the communicated information and will be considered in detail later.

Organizational concerns:

• Local traffic that originates at or terminates on nodes within an organization (also called autonomous system, AS)

• Transit traffic that passes through an AS

3.1.2 Standards of Information Quality

In text, the entropy per character depends on how many values the character can assume. Because a continuous signal can assume an infinite number of different value at a sample point, we are led to assume that a continuous signal must have an entropy of an infinite number of bits per sample.

This would be true if we required absolutely accurate reproduction of the continuous signal.

However, signals are transmitted to be heard, seen, or sensed. Only a certain degree of fidelity of reproduction is required. Thus, in dealing with the samples which specify continuous signals, Shannon introduces fidelity criterion. To reproduce the signal in a way meeting the fidelity criterion requires only a finite number of binary digits per sample per second, and hence we say that, within the accuracy imposed by a particular fidelity criterion, the entropy of a continuous source has a particular value in bits per sample or bits per second.

Standards of information quality help perform ordering of information bits by importance (to the user).

Man best handles information if encoded to his abilities. (Pierce, p.234)

In some cases, we can apply common sense in deciding user’s servicing quality needs. For example, in applications such as voice and video, users are somewhat tolerable of information loss, but very sensitive to delays. Conversely, in file transfer or electronic mail applications, the users are expected to be intolerable to loss and tolerable to delays. Finally, there are applications where both delay and loss can be aggravating to the user, such as in the case of interactive graphics or interactive computing applications.

For video, expectations are low

For voice, ear is very sensitive to jitter and latencies, and loss/flicker

Voice communication requires a steady, predictable packet delivery rate in order to maintain quality. Jitter, which is variation in packet delivery timing, is the most common culprit that reduces call quality in Internet telephony systems. Jitter causes the audio stream to become

broken, uneven or irregular. As a result, the listener’s experience becomes unpleasant or intolerable.

The end results of packet loss are similar to those of jitter but are typically more severe when the rate of packet loss is high. Excessive latency can result in unnatural conversation flow where there is a delay between words that one speaks versus words that one hears. Latency can cause callers to talk over one another and can also result in echoes on the line. Hence, jitter, packet loss and latency can have dramatic consequences in maintaining normal and expected call quality.

Human users are not the only recipients of information. For example, network management system exchanges signaling packets that may never reach human user.

These packets normally receive preferential treatment at the intermediaries (routers), and this is particularly required during times of congestion or failure.

It is particularly important during periods of congestion that traffic flows with different requirements be differentiated for servicing treatments. For example, a router might transmit higher-priority packets ahead of lower-priority packets in the same queue. Or a router may maintain different queues for different packet priorities and provide preferential treatment to the higher priority queues.

Dans le document Ivan Marsic (Page 193-197)