Next, the thesis will present and analyze the model of Nallapati et al. [128], which is used as a base model to develop the proposed summary model. Then, two mechanisms used in [43] will be presented to overcome the weaknesses of the base model [128]. Finally, the thesis proposes a solution to overcome the weaknesses of the model in [43] to improve the effectiveness of the summary model.
4.2. Basic summary model
Main components of the basic summarization-oriented single-text summarization model
[128] is shown in Figure 4.1 below.
Figure 4.1. Basic summarization-oriented single-text summarization model [128]
4.2.1. Seq2seq model of the model
In the seq2seq model, the encoding stage reads the input text
x x 1 , x 2 , x 3 ,..., x i ,..., x J and encode into encoded hidden states
1 2 3
j J
heh e, he, he, ..., he, ..., he , these encoded hidden states will be the input to the
decoder to generate output summary
y y 1 , y 2, y 3 ,..., y j,..., y T , in which:
- 𝑥 𝑖 , 𝑦 𝑗 are vectors of words from the input text and the summary text, respectively.
- 𝐽 , 𝑇 are the number of words of the input text and the summary text respectively.
The seq2seq model of the base model [128] with the encoder using a biLSTM network and the decoder using a LSTM network. The biLSTM network of the encoder encodes the sequence of input words x into the encoded hidden states
1 2 3
j J
jjj
j
j
heh e, he, he, ..., he, ..., heand
h e he he ; with:
he , he
dependent representation
1
J
in the forward and backward directions respectively, is the concatenation, the superscript e denotes that it is the encoding stage). The LSTM network of the encoder-decoder will take the
J
encoded representation of the text (hidden and remembered states)
he ,
he ,
c e ,
1
c e ) as input to generate the summary y . The encoded vectors are used to initialize the hidden states and memory states for the decoder as follows:
h d tanh( W ( h e h e ) b (4.3)
0
c d c e c e
e 2 d J 1
e 2 d
(4.4)
0 J 1
t
with: the superscript d denotes that it is the decoding stage, W e2d , 𝑏 e2d are the parameters
arithmetic. At each decoding step t , the hidden state
h d is updated based on hidden state
before
d
h
t 1
and input words
E (in the training phase it is the words in
y
t 1
The reference summary text is represented as:
tt 1
y t 1
h dLSTM h d, E
(4.5)
Then, the vocabulary distribution will be calculated according to the formula:
P vocab , t
softmax ( W d 2 v h d +b d 2 v )
(4.6)
t
with: W d2v , 𝑏 𝑑 2 𝑣 are learning parameters ,
P vocab , t
is a vector of size equal to size
of the vocabulary set 𝑉 . We will denote the probability of generating a target word w in the vocabulary set 𝑉 as
P vocab , t ( w ) .
4.2.2. Attention mechanism applied in the model
1 2 3
J
jjj
When in seq2seq model using attention mechanism, the decoder not only takes the encoded representations (hidden states and memory states) but also selects the key parts of the input at each decoding pass.
With the encoded hidden states being
heh e, he, he, ...., he
(with
h e he h e ). At
At each decoding step t , the attention score of each word is computed based on the hidden code states.
h
h
j
t
The e -encoded and d -decoded hidden states are as follows:
s es h e, h d ( v a lign ) Ttanh Wal lign heh d three lignes
(4.7)
tj jt
with: W align , v align , b align
jt
are the learning parameters.
Then, the attention distribution at this step t is computed based on the attention scores of all
all words in the input text
se , se , se , .... , se
by formula:
e
exp s e
exp s
, j 1 ,J
t 1 t 2 t 3 tJ
(4.8)
tj
tj J
i tk
k 1
The context vector of each word is computed using the following attention distribution:
J
t t j j
c e e h e
j 1
(4.9)
t
With the current hidden state:
h d , the hidden attention states are computed by the formula
h ˆ dW c eh d b
(4.10)
tcttc
Finally, the vocabulary distribution is calculated according to the formula:
t
P vocab , t softmax W d 2 v h ˆ d b d 2 v
where: W d2v , W c , b c and 𝑏 𝑑 2 𝑣 are learning parameters.
(4.11)
company
t
For t > 1, the hidden state
d
h
t 1
updated according to the formula:
h
d
t 1
LSTM h d, E h ˆ d
(4.12)
in there:
E h ˆ d
y
t
t
is input.
4.2.3. Word-copy network
Pointer-Generator networks allow both copying words in the source text and generating words in the vocabulary. This mechanism overcomes the problem of words not in the vocabulary. This mechanism does not affect the calculation of attention distribution and context vector, so attention distribution and context vector are still calculated according to formula (4.8) and formula (4.9) respectively. Context vector
t
h
t
c e , hidden decryption state d
and the decoder input E t will be the input to calculate the probability
𝑝 𝑔𝑒𝑛 (generation probability) according to the following formula:
p gene , t
( W c e W h d WE
b )
(4.13)
in there:
s , cts , hts , E ts
W s , c , W s , h , W s , E , b s
are the learning parameters.
( a ) is the sigmoid function.
p gene , t
is a real number and
p gene , t [0;1] .
p gene , t
is considered as a switch gate that can generate a word in the memory.
vocabulary or copy words from the source text using distributed attention (depending on whether the word is in the dictionary or not).
Call
P ( y t ) is the final distribution to predict a word, we have:
P ( y t )
p gene , tP g (y t ) 1 p gene , t P c (y t )
(4.14)
Vocabulary distribution
P g ( y t ) and attention distribution
P c ( y t ) is defined as follows:
P ( y ) P vocab , t ( y t ), y t V
(4.15)
gt 0 , y V
P ( y )
j : x j medical
t
, yes
e
tj t
V 1
(4.16)
ct
0 , y t V 1
with: V is the dictionary, V 1 are the words in the source text.
4.2.4. Coverage mechanism
The coverage mechanism proposed by Tu et al. [133] was initially used for the problem of machine translation using neural networks (NMT) to overcome the disadvantage of the attention mechanism which is to eliminate pre-existing information and avoid repetition. With this advantage, the coverage mechanism is applied to the problem of summarization.
u
t
text to solve the problem of word duplication. In this model, the covering vector e
is defined as the sum of the attentional distributions of the previous decoding steps, calculated by the formula:
t 1
t tj
u e e
j 1
(4.17)
Therefore, the coverage vector contains attention information on each word in the input text in the previous decoding steps. This coverage vector is used to recompute
Points of attention according to the following formula:
s e ( v a lign ) T tanh W a lign h e h d u e three lignes
(4.18)
tj jtt
The reason for this change in calculation is that word repetition problems can occur due to the solving step.
The current code depends too much on the previous decoding step, so we give the coverage vector (which is calculated by the sum of the attention distributions of the previous decoding steps) as input to solve this problem.
Then, we will calculate the coverage loss value according to the formula:
tj tj
j
covloss t
min( e , u e )
(4.19)
The coverage loss value will be used together with the hyperparameter 𝜆 and the loss function is calculated as:
t
loss t log P w *
min( e , u e )
(4.20)
tj tj
j
However, during the training phase the loss function is only calculated according to the following formula:
t
loss t log P w *
(4.21)
In the testing phase, the coverage loss function will be calculated as formula (4.19). Using the coverage loss function in the testing phase will help reduce the model training time [43].
4.3. PG_Feature_ASDS summary-oriented single-text summarization model
Pascanu et al. [134] pointed out that the weakness of the summarization models developed based on RNN networks is the gradient disappearance problem, that is: When the input text is too long, the first part of the text will be forgotten. The LSTM network model does not completely solve this problem. Since the main content of articles is often located at the beginning, See et al. [43] overcome this problem by taking only the first part of the article to feed into the model. However, this solution reduces the flexibility of
summarization model because not all types of documents have important content located at the beginning of the text. To address this problem, the model proposes to add position-of-sentence information (POSI) as a feature to the model to increase the weight of sentences at the beginning of the text without truncating the length of the input text.
In addition, the output word is generated based on the attention distribution of all input words on the encoder side and the previous output words. In addition, since we take the entire text without cutting off the end of the text, as the text size increases, the attention level of each word will decrease, so the attention efficiency will decrease. To overcome this problem, the proposed model uses additional word frequency (TF) features to help the model focus on important words.
The proposed new features for the model are detailed below.
4.3.1. New proposed features for the model
4.3.1.1. Sentence position characteristics
With input text
x x 1 , x 2 , x 3 ,...., x J
With k sentences, we can rewrite the vector as
x x 11 , x 21 , x 31 ,...., x Jk; in there:
x jk represents the 𝑗th word in the 𝑘th sentence . From vector 𝑥
With this , we can define a vector of length equal to vector x representing the position of the sentence.
contains the word:
x POSI
1,1,1,..., k , k .
Since the information is concentrated at the end of the text, we will increase the weight of the words in the text.
the top of the text up. Therefore,
x POSI is used to recalculate the attention score according to
formula (4.22) below and distribute attention according to formula (4.8) above.
( v align ) T t he W align h e h d b align
s
e
tjx
jt
POSI
(4.22)
4.3.1.2. Frequency characteristics of word occurrence
The frequency of words in a document is a parameter that determines the important words in a document. In a document, the higher the frequency of words, the higher the probability that the word is important. Using additional TF features to increase the weight of important words. The TF feature of a word is calculated according to the formula:
TF ( x i , x )
in there:
f ( x i , x )
max{f ( x i , x ) | j 1 J }
(4.23)
- 𝑓 ( 𝑥 𝑖 , 𝑥 ) is the number of times 𝑥 𝑖 appears in the text.
- max{ 𝑓 ( 𝑥 𝑗 , 𝑥 )| 𝑗 =1→ 𝐽 } is the maximum number of times a word appears in the text.
From vector 𝑥 , we can determine a vector with length equal to vector x representing the TF characteristic as follows:
x TF ( TF ( x 1 , x ), TF ( x 2 , x ), TF ( x 3 , x ),...., TF ( x J , x ))
(4.24)
We use vectors
x TF
This is to recalculate to recalculate the attention point according to the formula
(4.25) below and distribute attention according to formula (4.8) above.
s
jt
e
( v align ) T t he W align h e h d b align . x TF
tjx
(4.25)
POSI
4.3.2. Proposed summary-oriented single-text summarization model
The proposed model architecture includes a seq2seq model with an encoder using a biLSTM network and a decoder using an LSTM network, and an attention mechanism is used to help the model focus on the main information of the text. Although the seq2seq model uses an attention mechanism, it still has disadvantages such as word repetition, sentence repetition, and information loss. Therefore, the proposed model uses 2 mechanisms in [43] to solve the above problems:
- Coverage mechanism: Fixes word and sentence repetition errors.
- Copy-from mechanism: Corrects information loss.
However, during the summary testing process for English (CNN/Daily Mail dataset) and Vietnamese (Baomoi dataset), the model did not give the expected results, many test samples gave inaccurate results, so the thesis proposed to add 2 new text features to the model: Sentence position feature in text (POSI) and word frequency in text (TF).
The proposed model with newly added POSI and TF features is shown in Figure 4.2 below.
Figure 4.2. Proposed summary-oriented single-text summarization model PG_Feature_ASDS
4.4. Model testing
4.4.1. Test data sets
The proposed model is tested on two datasets CNN/Daily Mail for English and Baomoi for Vietnamese. The purpose of testing on the datasets
CNN/Daily Mail is to compare the results of the proposed model with the results of summary-oriented text summarization systems for English on the same recent dataset. The experiment on the Baomoi dataset is to evaluate the effectiveness of the proposed model for another language, Vietnamese, and to ensure the generality of the proposed summary-oriented summarization approach.
4.4.2. Data preprocessing
First, the input text dataset is processed by word separation using the Stanford CoreNLP library for English texts and the UETSegment 14 library for Vietnamese texts. For the texts of the Baomoi dataset, remove meaningless words in many texts (for example: vov.vn, dantri.vn, baodautu.vn, ...) because these words do not contribute to the content of the text, remove texts without a summary or without content, and articles that are too short (less than 50 characters) are also eliminated.
Then, each data unit (including 1 summary and 1 content) is formatted according to the data type specified in Tensorflow (for both datasets). This data type is formatted for all 3 datasets: Training dataset, Validation dataset and Test dataset. At the same time, based on the training data, a vocabulary set with a size of 50,000 words is created.
4.4.3. Experimental design
The thesis tested four different models on the CNN/Daily Mail and Baomoi datasets as follows:
(i) Model 1: Basic seq2seq model with attention mechanism [128].
(ii) Model 2: Pointer - Generator Network with Coverage mechanism [43].
(iii) Model 3: The proposed system is based on [43] and adds sentence position feature.
(iv) Model 4: The proposed system is based on [43] and adds sentence position and word frequency features.
Models 1 and 2 are tested by the source code in [43] on two datasets CNN/Daily Mail and Baomoi. Models 3 and 4 are implemented by the thesis to select the proposed summary model.
The input to the model is a sequence of words from the article, each word is represented as a vector. The vocabulary size in the experiments is 50,000 words for English and Vietnamese. For the experiments, the model has a 256-dimensional hidden state and a 128-dimensional word encoding vector, the batch size is limited to 16, and the input text length is 800 words for English and 550 words for Vietnamese (since English texts are less than 800 words long and Vietnamese texts are less than 550 words long, the text length is limited to this extent, which is reasonable). The model uses the Adagrad optimizer [135] with a learning factor of 0.15 and an initial accumulation value of 0.1. When fine-tuning the model, the loss function value is used to stop the model early. During the evaluation phase, the summary length was limited to a maximum of 100 words for both datasets.
In addition, the system also tested the model of See et al. [43] on the CNN/Daily Mail dataset to evaluate the effectiveness of using the first 400 words of the text as input for the system.
14 https://github.com/phongnt570/UETsegmenter
4.5. Evaluation and comparison of results
Table 4.1 below shows the experimental results on the CNN/Daily Mail dataset. R-1, R-2 and RL measures are used to evaluate and compare the performance of the models.
Model
CNN/Daily Mail | |||
R-1 | R-2 | RL | |
Model 1 (Seq2seq + attention) [128] | 27.21 | 10.09 | 24.48 |
Model 2 (Pointer-Generator + Coverage) [43] (*) | 29.71 | 12.13 | 28.05 |
Model 3 ( (*) + POSI) | 31.16 | 12.66 | 28.61 |
Model 4 ( (*) + POSI + TF) | 31.89 | 13.01 | 29.97 |
Maybe you are interested!
-
![Qos Assurance Methods for Multimedia Communications
zt2i3t4l5ee
zt2a3gs
zt2a3ge
zc2o3n4t5e6n7ts
low. The EF PHB requires a sufficiently large number of output ports to provide low delay, low loss, and low jitter.
EF PHBs can be implemented if the output ports bandwidth is sufficiently large, combined with small buffer sizes and other network resources dedicated to EF packets, to allow the routers service rate for EF packets on an output port to exceed the arrival rate λ of packets at that port.
This means that packets with PHB EF are considered with a pre-allocated amount of output bandwidth and a priority that ensures minimum loss, minimum delay and minimum jitter before being put into operation.
PHB EF is suitable for channel simulation, leased line simulation, and real-time services such as voice, video without compromising on high loss, delay and jitter values.
Figure 2.10 Example of EF installation
Figure 2.10 shows an example of an EF PHB implementation. This is a simple priority queue scheduling technique. At the edges of the DS domain, EF packet traffic is prioritized according to the values agreed upon by the SLA. The EF queue in the figure needs to output packets at a rate higher than the packet arrival rate λ. To provide an EF PHB over an end-to-end DS domain, bandwidth at the output ports of the core routers needs to be allocated in advance to ensure the requirement μ > λ. This can be done by a pre-configured provisioning process. In the figure, EF packets are placed in the priority queue (the upper queue). With such a length, the queue can operate with μ > λ.
Since EF was primarily used for real-time services such as voice and video, and since real-time services use UDP instead of TCP, RED is generally
not suitable for EF queues because applications using UDP will not respond to random packet drop and RED will strip unnecessary packets.
2.2.4.2 Assured Forwarding (AF) PHB
PHB AF is defined by RFC 2597. The purpose of PHB AF is to deliver packets reliably and therefore delay and jitter are considered less important than packet loss. PHB AF is suitable for non-real-time services such as applications using TCP. PHB AF first defines four classes: AF1, AF2, AF3, AF4. For each of these AF classes, packets are then classified into three subclasses with three distinct priority levels.
Table 2.8 shows the four AF classes and 12 AF subclasses and the DSCP values for the 12 AF subclasses defined by RFC 2597. RFC 2597 also allows for more than three separate priority levels to be added for internal use. However, these separate priority levels will only have internal significance.
PHB Class
PHB Subclass
Package type
DSCP
AF4
AF41
Short
100010
AF42
Medium
100100
AF43
High
100110
AF3
AF31
Short
011010
AF32
Medium
011100
AF33
High
011110
AF2
AF21
Short
010010
AF22
Medium
010100
AF23
High
010110
AF1
AF11
Short
001010
AF12
Medium
001100
AF13
High
001110
Table 2.8 AF DSCPs
The AF PHB ensures that packets are forwarded with a high probability of delivery to the destination within the bounds of the rate agreed upon in an SLA. If AF traffic at an ingress port exceeds the pre-priority rate, which is considered non-compliant or “out of profile”, the excess packets will not be delivered to the destination with the same probability as the packets belonging to the defined traffic or “in profile” packets. When there is network congestion, the out of profile packets are dropped before the in profile packets are dropped.
When service levels are defined using AF classes, different quantity and quality between AF classes can be realized by allocating different amounts of bandwidth and buffer space to the four AF classes. Unlike
EF, most AF traffic is non-real-time traffic using TCP, and the RED queue management strategy is an AQM (Adaptive Queue Management) strategy suitable for use in AF PHBs. The four AF PHB layers can be implemented as four separate queues. The output port bandwidth is divided into four AF queues. For each AF queue, packets are marked with three “colors” corresponding to three separate priority levels.
In addition to the 32 DSCP 1 groups defined in Table 2.8, 21 DSCPs have been standardized as follows: one for PHB EF, 12 for PHB AF, and 8 for CSCP. There are 11 DSCP 1 groups still available for other standards.
2.2.5.Example of Differentiated Services
We will look at an example of the Differentiated Service model and mechanism of operation. The architecture of Differentiated Service consists of two basic sets of functions:
Edge functions: include packet classification and traffic conditioning. At the inbound edge of the network, incoming packets are marked. In particular, the DS field in the packet header is set to a certain value. For example, in Figure 2.12, packets sent from H1 to H3 are marked at R1, while packets from H2 to H4 are marked at R2. The labels on the received packets identify the service class to which they belong. Different traffic classes receive different services in the core network. The RFC definition uses the term behavior aggregate rather than the term traffic class. After being marked, a packet can be forwarded immediately into the network, delayed for a period of time before being forwarded, or dropped. We will see that there are many factors that affect how a packet is marked, and whether it is forwarded immediately, delayed, or dropped.
Figure 2.12 DiffServ Example
Core functionality: When a DS-marked packet arrives at a Diffservcapable router, the packet is forwarded to the next router based on
Per-hop behavior is associated with packet classes. Per-hop behavior affects router buffers and the bandwidth shared between competing classes. An important principle of the Differentiated Service architecture is that a routers per-hop behavior is based only on the packets marking or the class to which it belongs. Therefore, if packets sent from H1 to H3 as shown in the figure receive the same marking as packets from H2 to H4, then the network routers treat the packets exactly the same, regardless of whether the packet originated from H1 or H2. For example, R3 does not distinguish between packets from h1 and H2 when forwarding packets to R4. Therefore, the Differentiated Service architecture avoids the need to maintain router state about separate source-destination pairs, which is important for network scalability.
Chapter Conclusion
Chapter 2 has presented and clarified two main models of deploying and installing quality of service in IP networks. While the traditional best-effort model has many disadvantages, later models such as IntServ and DiffServ have partly solved the problems that best-effort could not solve. IntServ follows the direction of ensuring quality of service for each separate flow, it is built similar to the circuit switching model with the use of the RSVP resource reservation protocol. IntSer is suitable for services that require fixed bandwidth that is not shared such as VoIP services, multicast TV services. However, IntSer has disadvantages such as using a lot of network resources, low scalability and lack of flexibility. DiffServ was born with the idea of solving the disadvantages of the IntServ model.
DiffServ follows the direction of ensuring quality based on the principle of hop-by-hop behavior based on the priority of marked packets. The policy for different types of traffic is decided by the administrator and can be changed according to reality, so it is very flexible. DiffServ makes better use of network resources, avoiding idle bandwidth and processing capacity on routers. In addition, the DifServ model can be deployed on many independent domains, so the ability to expand the network becomes easy.
Chapter 3: METHODS TO ENSURE QoS FOR MULTIMEDIA COMMUNICATIONS
In packet-switched networks, different packet flows often have to share the transmission medium all the way to the destination station. To ensure the fair and efficient allocation of bandwidth to flows, appropriate serving mechanisms are required at network nodes, especially at gateways or routers, where many different data flows often pass through. The scheduler is responsible for serving packets of the selected flow and deciding which packet will be served next. Here, a flow is understood as a set of packets belonging to the same priority class, or originating from the same source, or having the same source and destination addresses, etc.
In normal state when there is no congestion, packets will be sent as soon as they are delivered. In case of congestion, if QoS assurance methods are not applied, prolonged congestion can cause packet drops, affecting service quality. In some cases, congestion is prolonged and widespread in the network, which can easily lead to the network being frozen, or many packets being dropped, seriously affecting service quality.
Therefore, in this chapter, in sections 3.2 and 3.3, we introduce some typical network traffic load monitoring techniques to predict and prevent congestion before it occurs through the measure of dropping (removing) packets early when there are signs of impending congestion.
3.1. DropTail method
DropTail is a simple, traditional queue management method based on FIFO mechanism. All incoming packets are placed in the queue, when the queue is full, the later packets are dropped.
Due to its simplicity and ease of implementation, DropTail has been used for many years on Internet router systems. However, this algorithm has the following disadvantages:
− Cannot avoid the phenomenon of “Lock out”: Occurs when 1 or several traffic streams monopolize the queue, making packets of other connections unable to pass through the router. This phenomenon greatly affects reliable transmission protocols such as TCP. According to the anti-congestion algorithm, when locked out, the TCP connection stream will reduce the window size and reduce the packet transmission speed exponentially.
− Can cause Global Synchronization: This is the result of a severe “Lock out” phenomenon. Some neighboring routers have their queues monopolized by a number of connections, causing a series of other TCP connections to be unable to pass through and simultaneously reducing the transmission speed. After those monopolized connections are temporarily suspended,
Once the queue is cleared, it takes a considerable amount of time for TCP connections to return to their original speed.
− Full Queue phenomenon: Data transmitted on the Internet often has an explosion, packets arriving at the router are often in clusters rather than in turn. Therefore, the operating mechanism of DropTail makes the queue easily full for a long period of time, leading to the average delay time of large packets. To avoid this phenomenon, with DropTail, the only way is to increase the routers buffer, this method is very expensive and ineffective.
− No QoS guarantee: With the DropTail mechanism, there is no way to prioritize important packets to be transmitted through the router earlier when all are in the queue. Meanwhile, with multimedia communication, ensuring connection and stable speed is extremely important and the DropTail algorithm cannot satisfy.
The problem of choosing the buffer size of the routers in the network is to “absorb” short bursts of traffic without causing too much queuing delay. This is necessary in bursty data transmission. The queue size determines the size of the packet bursts (traffic spikes) that we want to be able to transmit without being dropped at the routers.
In IP-based application networks, packet dropping is an important mechanism for indirectly reporting congestion to end stations. A solution that prevents router queues from filling up while reducing the packet drop rate is called dynamic queue management.
3.2. Random elimination method – RED
3.2.1 Overview
RED (Random Early Detection of congestion; Random Early Drop) is one of the first AQM algorithms proposed in 1993 by Sally Floyd and Van Jacobson, two scientists at the Lawrence Berkeley Laboratory of the University of California, USA. Due to its outstanding advantages compared to previous queue management algorithms, RED has been widely installed and deployed on the Internet.
The most fundamental point of their work is that the most effective place to detect congestion and react to it is at the gateway or router.
Source entities (senders) can also do this by estimating end-to-end delay, throughput variability, or the rate of packet retransmissions due to drop. However, the sender and receiver view of a particular connection cannot tell which gateways on the network are congested, and cannot distinguish between propagation delay and queuing delay. Only the gateway has a true view of the state of the queue, the link share of the connections passing through it at any given time, and the quality of service requirements of the
traffic flows. The RED gateway monitors the average queue length, which detects early signs of impending congestion (average queue length exceeding a predetermined threshold) and reacts appropriately in one of two ways:
− Drop incoming packets with a certain probability, to indirectly inform the source of congestion, the source needs to reduce the transmission rate to keep the queue from filling up, maintaining the ability to absorb incoming traffic spikes.
− Mark “congestion” with a certain probability in the ECN field in the header of TCP packets to notify the source (the receiving entity will copy this bit into the acknowledgement packet).
Figure 3. 1 RED algorithm
The main goal of RED is to avoid congestion by keeping the average queue size within a sufficiently small and stable region, which also means keeping the queuing delay sufficiently small and stable. Achieving this goal also helps: avoid global synchronization, not resist bursty traffic flows (i.e. flows with low average throughput but high volatility), and maintain an upper bound on the average queue size even in the absence of cooperation from transport layer protocols.
To achieve the above goals, RED gateways must do the following:
− The first is to detect congestion early and react appropriately to keep the average queue size small enough to keep the network operating in the low latency, high throughput region, while still allowing the queue size to fluctuate within a certain range to absorb short-term fluctuations. As discussed above, the gateway is the most appropriate place to detect congestion and is also the most appropriate place to decide which specific connection to report congestion to.
− The second thing is to notify the source of congestion. This is done by marking and notifying the source to reduce traffic. Normally the RED gateway will randomly drop packets. However, if congestion
If congestion is detected before the queue is full, it should be combined with packet marking to signal congestion. The RED gateway has two options: drop or mark; where marking is done by marking the ECN field of the packet with a certain probability, to signal the source to reduce the traffic entering the network.
− An important goal that RED gateways need to achieve is to avoid global synchronization and not to resist traffic flows that have a sudden characteristic. Global synchronization occurs when all connections simultaneously reduce their transmission window size, leading to a severe drop in throughput at the same time. On the other hand, Drop Tail or Random Drop strategies are very sensitive to sudden flows; that is, the gateway queue will often overflow when packets from these flows arrive. To avoid these two phenomena, gateways can use special algorithms to detect congestion and decide which connections will be notified of congestion at the gateway. The RED gateway randomly selects incoming packets to mark; with this method, the probability of marking a packet from a particular connection is proportional to the connections shared bandwidth at the gateway.
− Another goal is to control the average queue size even without cooperation from the source entities. This can be done by dropping packets when the average size exceeds an upper threshold (instead of marking it). This approach is necessary in cases where most connections have transmission times that are less than the round-trip time, or where the source entities are not able to reduce traffic in response to marking or dropping packets (such as UDP flows).
3.2.2 Algorithm
This section describes the algorithm for RED gateways. RED gateways calculate the average queue size using a low-pass filter. This average queue size is compared with two thresholds: minth and maxth. When the average queue size is less than the lower threshold, no incoming packets are marked or dropped; when the average queue size is greater than the upper threshold, all incoming packets are dropped. When the average queue size is between minth and maxth, each incoming packet is marked or dropped with a probability pa, where pa is a function of the average queue size avg; the probability of marking or dropping a packet for a particular connection is proportional to the bandwidth share of that connection at the gateway. The general algorithm for a RED gateway is described as follows: [5]
For each packet arrival
Caculate the average queue size avg If minth ≤ avg < maxth
div.maincontent .s1 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 15pt; }
div.maincontent .s2 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: bold; text-decoration: none; font-size: 15pt; }
div.maincontent .p { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; margin:0pt; }
div.maincontent p { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; margin:0pt; }
div.maincontent .s3 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: bold; text-decoration: none; font-size: 14pt; }
div.maincontent .s4 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; }
div.maincontent .s5 { color: black; font-family:Times New Roman, serif; font-style: italic; font-weight: normal; text-decoration: none; font-size: 14pt; }
div.maincontent .s6 { color: black; font-family:Times New Roman, serif; font-style: italic; font-weight: bold; text-decoration: none; font-size: 14pt; }
div.maincontent .s7 { color: black; font-family:Wingdings; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; }
div.maincontent .s8 { color: black; font-family:Arial, sans-serif; font-style: italic; font-weight: bold; text-decoration: none; font-size: 15pt; }
div.maincontent .s9 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: bold; text-decoration: none; font-size: 14pt; }
div.maincontent .s10 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 9pt; vertical-align: 6pt; }
div.maincontent .s11 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 13pt; }
div.maincontent .s12 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 10pt; }
div.maincontent .s13 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-d](https://tailieuthamkhao.com/uploads/2022/05/15/danh-gia-hieu-qua-dam-bao-qos-cho-truyen-thong-da-phuong-tien-cua-chien-6-1-120x90.jpg)
![Qos Assurance Methods for Multimedia Communications
zt2i3t4l5ee
zt2a3gs
zt2a3ge
zc2o3n4t5e6n7ts
low. The EF PHB requires a sufficiently large number of output ports to provide low delay, low loss, and low jitter.
EF PHBs can be implemented if the output ports bandwidth is sufficiently large, combined with small buffer sizes and other network resources dedicated to EF packets, to allow the routers service rate for EF packets on an output port to exceed the arrival rate λ of packets at that port.
This means that packets with PHB EF are considered with a pre-allocated amount of output bandwidth and a priority that ensures minimum loss, minimum delay and minimum jitter before being put into operation.
PHB EF is suitable for channel simulation, leased line simulation, and real-time services such as voice, video without compromising on high loss, delay and jitter values.
Figure 2.10 Example of EF installation
Figure 2.10 shows an example of an EF PHB implementation. This is a simple priority queue scheduling technique. At the edges of the DS domain, EF packet traffic is prioritized according to the values agreed upon by the SLA. The EF queue in the figure needs to output packets at a rate higher than the packet arrival rate λ. To provide an EF PHB over an end-to-end DS domain, bandwidth at the output ports of the core routers needs to be allocated in advance to ensure the requirement μ > λ. This can be done by a pre-configured provisioning process. In the figure, EF packets are placed in the priority queue (the upper queue). With such a length, the queue can operate with μ > λ.
Since EF was primarily used for real-time services such as voice and video, and since real-time services use UDP instead of TCP, RED is generally
not suitable for EF queues because applications using UDP will not respond to random packet drop and RED will strip unnecessary packets.
2.2.4.2 Assured Forwarding (AF) PHB
PHB AF is defined by RFC 2597. The purpose of PHB AF is to deliver packets reliably and therefore delay and jitter are considered less important than packet loss. PHB AF is suitable for non-real-time services such as applications using TCP. PHB AF first defines four classes: AF1, AF2, AF3, AF4. For each of these AF classes, packets are then classified into three subclasses with three distinct priority levels.
Table 2.8 shows the four AF classes and 12 AF subclasses and the DSCP values for the 12 AF subclasses defined by RFC 2597. RFC 2597 also allows for more than three separate priority levels to be added for internal use. However, these separate priority levels will only have internal significance.
PHB Class
PHB Subclass
Package type
DSCP
AF4
AF41
Short
100010
AF42
Medium
100100
AF43
High
100110
AF3
AF31
Short
011010
AF32
Medium
011100
AF33
High
011110
AF2
AF21
Short
010010
AF22
Medium
010100
AF23
High
010110
AF1
AF11
Short
001010
AF12
Medium
001100
AF13
High
001110
Table 2.8 AF DSCPs
The AF PHB ensures that packets are forwarded with a high probability of delivery to the destination within the bounds of the rate agreed upon in an SLA. If AF traffic at an ingress port exceeds the pre-priority rate, which is considered non-compliant or “out of profile”, the excess packets will not be delivered to the destination with the same probability as the packets belonging to the defined traffic or “in profile” packets. When there is network congestion, the out of profile packets are dropped before the in profile packets are dropped.
When service levels are defined using AF classes, different quantity and quality between AF classes can be realized by allocating different amounts of bandwidth and buffer space to the four AF classes. Unlike
EF, most AF traffic is non-real-time traffic using TCP, and the RED queue management strategy is an AQM (Adaptive Queue Management) strategy suitable for use in AF PHBs. The four AF PHB layers can be implemented as four separate queues. The output port bandwidth is divided into four AF queues. For each AF queue, packets are marked with three “colors” corresponding to three separate priority levels.
In addition to the 32 DSCP 1 groups defined in Table 2.8, 21 DSCPs have been standardized as follows: one for PHB EF, 12 for PHB AF, and 8 for CSCP. There are 11 DSCP 1 groups still available for other standards.
2.2.5.Example of Differentiated Services
We will look at an example of the Differentiated Service model and mechanism of operation. The architecture of Differentiated Service consists of two basic sets of functions:
Edge functions: include packet classification and traffic conditioning. At the inbound edge of the network, incoming packets are marked. In particular, the DS field in the packet header is set to a certain value. For example, in Figure 2.12, packets sent from H1 to H3 are marked at R1, while packets from H2 to H4 are marked at R2. The labels on the received packets identify the service class to which they belong. Different traffic classes receive different services in the core network. The RFC definition uses the term behavior aggregate rather than the term traffic class. After being marked, a packet can be forwarded immediately into the network, delayed for a period of time before being forwarded, or dropped. We will see that there are many factors that affect how a packet is marked, and whether it is forwarded immediately, delayed, or dropped.
Figure 2.12 DiffServ Example
Core functionality: When a DS-marked packet arrives at a Diffservcapable router, the packet is forwarded to the next router based on
Per-hop behavior is associated with packet classes. Per-hop behavior affects router buffers and the bandwidth shared between competing classes. An important principle of the Differentiated Service architecture is that a routers per-hop behavior is based only on the packets marking or the class to which it belongs. Therefore, if packets sent from H1 to H3 as shown in the figure receive the same marking as packets from H2 to H4, then the network routers treat the packets exactly the same, regardless of whether the packet originated from H1 or H2. For example, R3 does not distinguish between packets from h1 and H2 when forwarding packets to R4. Therefore, the Differentiated Service architecture avoids the need to maintain router state about separate source-destination pairs, which is important for network scalability.
Chapter Conclusion
Chapter 2 has presented and clarified two main models of deploying and installing quality of service in IP networks. While the traditional best-effort model has many disadvantages, later models such as IntServ and DiffServ have partly solved the problems that best-effort could not solve. IntServ follows the direction of ensuring quality of service for each separate flow, it is built similar to the circuit switching model with the use of the RSVP resource reservation protocol. IntSer is suitable for services that require fixed bandwidth that is not shared such as VoIP services, multicast TV services. However, IntSer has disadvantages such as using a lot of network resources, low scalability and lack of flexibility. DiffServ was born with the idea of solving the disadvantages of the IntServ model.
DiffServ follows the direction of ensuring quality based on the principle of hop-by-hop behavior based on the priority of marked packets. The policy for different types of traffic is decided by the administrator and can be changed according to reality, so it is very flexible. DiffServ makes better use of network resources, avoiding idle bandwidth and processing capacity on routers. In addition, the DifServ model can be deployed on many independent domains, so the ability to expand the network becomes easy.
Chapter 3: METHODS TO ENSURE QoS FOR MULTIMEDIA COMMUNICATIONS
In packet-switched networks, different packet flows often have to share the transmission medium all the way to the destination station. To ensure the fair and efficient allocation of bandwidth to flows, appropriate serving mechanisms are required at network nodes, especially at gateways or routers, where many different data flows often pass through. The scheduler is responsible for serving packets of the selected flow and deciding which packet will be served next. Here, a flow is understood as a set of packets belonging to the same priority class, or originating from the same source, or having the same source and destination addresses, etc.
In normal state when there is no congestion, packets will be sent as soon as they are delivered. In case of congestion, if QoS assurance methods are not applied, prolonged congestion can cause packet drops, affecting service quality. In some cases, congestion is prolonged and widespread in the network, which can easily lead to the network being frozen, or many packets being dropped, seriously affecting service quality.
Therefore, in this chapter, in sections 3.2 and 3.3, we introduce some typical network traffic load monitoring techniques to predict and prevent congestion before it occurs through the measure of dropping (removing) packets early when there are signs of impending congestion.
3.1. DropTail method
DropTail is a simple, traditional queue management method based on FIFO mechanism. All incoming packets are placed in the queue, when the queue is full, the later packets are dropped.
Due to its simplicity and ease of implementation, DropTail has been used for many years on Internet router systems. However, this algorithm has the following disadvantages:
− Cannot avoid the phenomenon of “Lock out”: Occurs when 1 or several traffic streams monopolize the queue, making packets of other connections unable to pass through the router. This phenomenon greatly affects reliable transmission protocols such as TCP. According to the anti-congestion algorithm, when locked out, the TCP connection stream will reduce the window size and reduce the packet transmission speed exponentially.
− Can cause Global Synchronization: This is the result of a severe “Lock out” phenomenon. Some neighboring routers have their queues monopolized by a number of connections, causing a series of other TCP connections to be unable to pass through and simultaneously reducing the transmission speed. After those monopolized connections are temporarily suspended,
Once the queue is cleared, it takes a considerable amount of time for TCP connections to return to their original speed.
− Full Queue phenomenon: Data transmitted on the Internet often has an explosion, packets arriving at the router are often in clusters rather than in turn. Therefore, the operating mechanism of DropTail makes the queue easily full for a long period of time, leading to the average delay time of large packets. To avoid this phenomenon, with DropTail, the only way is to increase the routers buffer, this method is very expensive and ineffective.
− No QoS guarantee: With the DropTail mechanism, there is no way to prioritize important packets to be transmitted through the router earlier when all are in the queue. Meanwhile, with multimedia communication, ensuring connection and stable speed is extremely important and the DropTail algorithm cannot satisfy.
The problem of choosing the buffer size of the routers in the network is to “absorb” short bursts of traffic without causing too much queuing delay. This is necessary in bursty data transmission. The queue size determines the size of the packet bursts (traffic spikes) that we want to be able to transmit without being dropped at the routers.
In IP-based application networks, packet dropping is an important mechanism for indirectly reporting congestion to end stations. A solution that prevents router queues from filling up while reducing the packet drop rate is called dynamic queue management.
3.2. Random elimination method – RED
3.2.1 Overview
RED (Random Early Detection of congestion; Random Early Drop) is one of the first AQM algorithms proposed in 1993 by Sally Floyd and Van Jacobson, two scientists at the Lawrence Berkeley Laboratory of the University of California, USA. Due to its outstanding advantages compared to previous queue management algorithms, RED has been widely installed and deployed on the Internet.
The most fundamental point of their work is that the most effective place to detect congestion and react to it is at the gateway or router.
Source entities (senders) can also do this by estimating end-to-end delay, throughput variability, or the rate of packet retransmissions due to drop. However, the sender and receiver view of a particular connection cannot tell which gateways on the network are congested, and cannot distinguish between propagation delay and queuing delay. Only the gateway has a true view of the state of the queue, the link share of the connections passing through it at any given time, and the quality of service requirements of the
traffic flows. The RED gateway monitors the average queue length, which detects early signs of impending congestion (average queue length exceeding a predetermined threshold) and reacts appropriately in one of two ways:
− Drop incoming packets with a certain probability, to indirectly inform the source of congestion, the source needs to reduce the transmission rate to keep the queue from filling up, maintaining the ability to absorb incoming traffic spikes.
− Mark “congestion” with a certain probability in the ECN field in the header of TCP packets to notify the source (the receiving entity will copy this bit into the acknowledgement packet).
Figure 3. 1 RED algorithm
The main goal of RED is to avoid congestion by keeping the average queue size within a sufficiently small and stable region, which also means keeping the queuing delay sufficiently small and stable. Achieving this goal also helps: avoid global synchronization, not resist bursty traffic flows (i.e. flows with low average throughput but high volatility), and maintain an upper bound on the average queue size even in the absence of cooperation from transport layer protocols.
To achieve the above goals, RED gateways must do the following:
− The first is to detect congestion early and react appropriately to keep the average queue size small enough to keep the network operating in the low latency, high throughput region, while still allowing the queue size to fluctuate within a certain range to absorb short-term fluctuations. As discussed above, the gateway is the most appropriate place to detect congestion and is also the most appropriate place to decide which specific connection to report congestion to.
− The second thing is to notify the source of congestion. This is done by marking and notifying the source to reduce traffic. Normally the RED gateway will randomly drop packets. However, if congestion
If congestion is detected before the queue is full, it should be combined with packet marking to signal congestion. The RED gateway has two options: drop or mark; where marking is done by marking the ECN field of the packet with a certain probability, to signal the source to reduce the traffic entering the network.
− An important goal that RED gateways need to achieve is to avoid global synchronization and not to resist traffic flows that have a sudden characteristic. Global synchronization occurs when all connections simultaneously reduce their transmission window size, leading to a severe drop in throughput at the same time. On the other hand, Drop Tail or Random Drop strategies are very sensitive to sudden flows; that is, the gateway queue will often overflow when packets from these flows arrive. To avoid these two phenomena, gateways can use special algorithms to detect congestion and decide which connections will be notified of congestion at the gateway. The RED gateway randomly selects incoming packets to mark; with this method, the probability of marking a packet from a particular connection is proportional to the connections shared bandwidth at the gateway.
− Another goal is to control the average queue size even without cooperation from the source entities. This can be done by dropping packets when the average size exceeds an upper threshold (instead of marking it). This approach is necessary in cases where most connections have transmission times that are less than the round-trip time, or where the source entities are not able to reduce traffic in response to marking or dropping packets (such as UDP flows).
3.2.2 Algorithm
This section describes the algorithm for RED gateways. RED gateways calculate the average queue size using a low-pass filter. This average queue size is compared with two thresholds: minth and maxth. When the average queue size is less than the lower threshold, no incoming packets are marked or dropped; when the average queue size is greater than the upper threshold, all incoming packets are dropped. When the average queue size is between minth and maxth, each incoming packet is marked or dropped with a probability pa, where pa is a function of the average queue size avg; the probability of marking or dropping a packet for a particular connection is proportional to the bandwidth share of that connection at the gateway. The general algorithm for a RED gateway is described as follows: [5]
For each packet arrival
Caculate the average queue size avg If minth ≤ avg < maxth
div.maincontent .s1 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 15pt; }
div.maincontent .s2 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: bold; text-decoration: none; font-size: 15pt; }
div.maincontent .p { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; margin:0pt; }
div.maincontent p { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; margin:0pt; }
div.maincontent .s3 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: bold; text-decoration: none; font-size: 14pt; }
div.maincontent .s4 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; }
div.maincontent .s5 { color: black; font-family:Times New Roman, serif; font-style: italic; font-weight: normal; text-decoration: none; font-size: 14pt; }
div.maincontent .s6 { color: black; font-family:Times New Roman, serif; font-style: italic; font-weight: bold; text-decoration: none; font-size: 14pt; }
div.maincontent .s7 { color: black; font-family:Wingdings; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; }
div.maincontent .s8 { color: black; font-family:Arial, sans-serif; font-style: italic; font-weight: bold; text-decoration: none; font-size: 15pt; }
div.maincontent .s9 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: bold; text-decoration: none; font-size: 14pt; }
div.maincontent .s10 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 9pt; vertical-align: 6pt; }
div.maincontent .s11 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 13pt; }
div.maincontent .s12 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 10pt; }
div.maincontent .s13 { color: black; font-family:Times New Roman, serif; font-style: normal; font-weight: normal; text-d](data:image/svg+xml,%3Csvg%20xmlns=%22http://www.w3.org/2000/svg%22%20viewBox=%220%200%2075%2075%22%3E%3C/svg%3E)
Qos Assurance Methods for Multimedia Communications
zt2i3t4l5ee
zt2a3gs
zt2a3ge
zc2o3n4t5e6n7ts
low. The EF PHB requires a sufficiently large number of output ports to provide low delay, low loss, and low jitter.
EF PHBs can be implemented if the output port's bandwidth is sufficiently large, combined with small buffer sizes and other network resources dedicated to EF packets, to allow the router's service rate for EF packets on an output port to exceed the arrival rate λ of packets at that port.
This means that packets with PHB EF are considered with a pre-allocated amount of output bandwidth and a priority that ensures minimum loss, minimum delay and minimum jitter before being put into operation.
PHB EF is suitable for channel simulation, leased line simulation, and real-time services such as voice, video without compromising on high loss, delay and jitter values.
Figure 2.10 Example of EF installation
Figure 2.10 shows an example of an EF PHB implementation. This is a simple priority queue scheduling technique. At the edges of the DS domain, EF packet traffic is prioritized according to the values agreed upon by the SLA. The EF queue in the figure needs to output packets at a rate higher than the packet arrival rate λ. To provide an EF PHB over an end-to-end DS domain, bandwidth at the output ports of the core routers needs to be allocated in advance to ensure the requirement μ > λ. This can be done by a pre-configured provisioning process. In the figure, EF packets are placed in the priority queue (the upper queue). With such a length, the queue can operate with μ > λ.
Since EF was primarily used for real-time services such as voice and video, and since real-time services use UDP instead of TCP, RED is generally
not suitable for EF queues because applications using UDP will not respond to random packet drop and RED will strip unnecessary packets.
2.2.4.2 Assured Forwarding (AF) PHB
PHB AF is defined by RFC 2597. The purpose of PHB AF is to deliver packets reliably and therefore delay and jitter are considered less important than packet loss. PHB AF is suitable for non-real-time services such as applications using TCP. PHB AF first defines four classes: AF1, AF2, AF3, AF4. For each of these AF classes, packets are then classified into three subclasses with three distinct priority levels.
Table 2.8 shows the four AF classes and 12 AF subclasses and the DSCP values for the 12 AF subclasses defined by RFC 2597. RFC 2597 also allows for more than three separate priority levels to be added for internal use. However, these separate priority levels will only have internal significance.
PHB Class
PHB Subclass
Package type
DSCP
AF4
AF41
Short
100010
AF42
Medium
100100
AF43
High
100110
AF3
AF31
Short
011010
AF32
Medium
011100
AF33
High
011110
AF2
AF21
Short
010010
AF22
Medium
010100
AF23
High
010110
AF1
AF11
Short
001010
AF12
Medium
001100
AF13
High
001110
Table 2.8 AF DSCPs
The AF PHB ensures that packets are forwarded with a high probability of delivery to the destination within the bounds of the rate agreed upon in an SLA. If AF traffic at an ingress port exceeds the pre-priority rate, which is considered non-compliant or “out of profile”, the excess packets will not be delivered to the destination with the same probability as the packets belonging to the defined traffic or “in profile” packets. When there is network congestion, the out of profile packets are dropped before the in profile packets are dropped.
When service levels are defined using AF classes, different quantity and quality between AF classes can be realized by allocating different amounts of bandwidth and buffer space to the four AF classes. Unlike
EF, most AF traffic is non-real-time traffic using TCP, and the RED queue management strategy is an AQM (Adaptive Queue Management) strategy suitable for use in AF PHBs. The four AF PHB layers can be implemented as four separate queues. The output port bandwidth is divided into four AF queues. For each AF queue, packets are marked with three “colors” corresponding to three separate priority levels.
In addition to the 32 DSCP 1 groups defined in Table 2.8, 21 DSCPs have been standardized as follows: one for PHB EF, 12 for PHB AF, and 8 for CSCP. There are 11 DSCP 1 groups still available for other standards.
2.2.5.Example of Differentiated Services
We will look at an example of the Differentiated Service model and mechanism of operation. The architecture of Differentiated Service consists of two basic sets of functions:
Edge functions: include packet classification and traffic conditioning. At the inbound edge of the network, incoming packets are marked. In particular, the DS field in the packet header is set to a certain value. For example, in Figure 2.12, packets sent from H1 to H3 are marked at R1, while packets from H2 to H4 are marked at R2. The labels on the received packets identify the service class to which they belong. Different traffic classes receive different services in the core network. The RFC definition uses the term behavior aggregate rather than the term traffic class. After being marked, a packet can be forwarded immediately into the network, delayed for a period of time before being forwarded, or dropped. We will see that there are many factors that affect how a packet is marked, and whether it is forwarded immediately, delayed, or dropped.
Figure 2.12 DiffServ Example
Core functionality: When a DS-marked packet arrives at a Diffservcapable router, the packet is forwarded to the next router based on
Per-hop behavior is associated with packet classes. Per-hop behavior affects router buffers and the bandwidth shared between competing classes. An important principle of the Differentiated Service architecture is that a router's per-hop behavior is based only on the packet's marking or the class to which it belongs. Therefore, if packets sent from H1 to H3 as shown in the figure receive the same marking as packets from H2 to H4, then the network routers treat the packets exactly the same, regardless of whether the packet originated from H1 or H2. For example, R3 does not distinguish between packets from h1 and H2 when forwarding packets to R4. Therefore, the Differentiated Service architecture avoids the need to maintain router state about separate source-destination pairs, which is important for network scalability.
Chapter Conclusion
Chapter 2 has presented and clarified two main models of deploying and installing quality of service in IP networks. While the traditional best-effort model has many disadvantages, later models such as IntServ and DiffServ have partly solved the problems that best-effort could not solve. IntServ follows the direction of ensuring quality of service for each separate flow, it is built similar to the circuit switching model with the use of the RSVP resource reservation protocol. IntSer is suitable for services that require fixed bandwidth that is not shared such as VoIP services, multicast TV services. However, IntSer has disadvantages such as using a lot of network resources, low scalability and lack of flexibility. DiffServ was born with the idea of solving the disadvantages of the IntServ model.
DiffServ follows the direction of ensuring quality based on the principle of hop-by-hop behavior based on the priority of marked packets. The policy for different types of traffic is decided by the administrator and can be changed according to reality, so it is very flexible. DiffServ makes better use of network resources, avoiding idle bandwidth and processing capacity on routers. In addition, the DifServ model can be deployed on many independent domains, so the ability to expand the network becomes easy.
Chapter 3: METHODS TO ENSURE QoS FOR MULTIMEDIA COMMUNICATIONS
In packet-switched networks, different packet flows often have to share the transmission medium all the way to the destination station. To ensure the fair and efficient allocation of bandwidth to flows, appropriate serving mechanisms are required at network nodes, especially at gateways or routers, where many different data flows often pass through. The scheduler is responsible for serving packets of the selected flow and deciding which packet will be served next. Here, a flow is understood as a set of packets belonging to the same priority class, or originating from the same source, or having the same source and destination addresses, etc.
In normal state when there is no congestion, packets will be sent as soon as they are delivered. In case of congestion, if QoS assurance methods are not applied, prolonged congestion can cause packet drops, affecting service quality. In some cases, congestion is prolonged and widespread in the network, which can easily lead to the network being "frozen", or many packets being dropped, seriously affecting service quality.
Therefore, in this chapter, in sections 3.2 and 3.3, we introduce some typical network traffic load monitoring techniques to predict and prevent congestion before it occurs through the measure of dropping (removing) packets early when there are signs of impending congestion.
3.1. DropTail method
DropTail is a simple, traditional queue management method based on FIFO mechanism. All incoming packets are placed in the queue, when the queue is full, the later packets are dropped.
Due to its simplicity and ease of implementation, DropTail has been used for many years on Internet router systems. However, this algorithm has the following disadvantages:
− Cannot avoid the phenomenon of “Lock out”: Occurs when 1 or several traffic streams monopolize the queue, making packets of other connections unable to pass through the router. This phenomenon greatly affects reliable transmission protocols such as TCP. According to the anti-congestion algorithm, when locked out, the TCP connection stream will reduce the window size and reduce the packet transmission speed exponentially.
− Can cause Global Synchronization: This is the result of a severe “Lock out” phenomenon. Some neighboring routers have their queues monopolized by a number of connections, causing a series of other TCP connections to be unable to pass through and simultaneously reducing the transmission speed. After those monopolized connections are temporarily suspended,
Once the queue is cleared, it takes a considerable amount of time for TCP connections to return to their original speed.
− Full Queue phenomenon: Data transmitted on the Internet often has an explosion, packets arriving at the router are often in clusters rather than in turn. Therefore, the operating mechanism of DropTail makes the queue easily full for a long period of time, leading to the average delay time of large packets. To avoid this phenomenon, with DropTail, the only way is to increase the router's buffer, this method is very expensive and ineffective.
− No QoS guarantee: With the DropTail mechanism, there is no way to prioritize important packets to be transmitted through the router earlier when all are in the queue. Meanwhile, with multimedia communication, ensuring connection and stable speed is extremely important and the DropTail algorithm cannot satisfy.
The problem of choosing the buffer size of the routers in the network is to “absorb” short bursts of traffic without causing too much queuing delay. This is necessary in bursty data transmission. The queue size determines the size of the packet bursts (traffic spikes) that we want to be able to transmit without being dropped at the routers.
In IP-based application networks, packet dropping is an important mechanism for indirectly reporting congestion to end stations. A solution that prevents router queues from filling up while reducing the packet drop rate is called dynamic queue management.
3.2. Random elimination method – RED
3.2.1 Overview
RED (Random Early Detection of congestion; Random Early Drop) is one of the first AQM algorithms proposed in 1993 by Sally Floyd and Van Jacobson, two scientists at the Lawrence Berkeley Laboratory of the University of California, USA. Due to its outstanding advantages compared to previous queue management algorithms, RED has been widely installed and deployed on the Internet.
The most fundamental point of their work is that the most effective place to detect congestion and react to it is at the gateway or router.
Source entities (senders) can also do this by estimating end-to-end delay, throughput variability, or the rate of packet retransmissions due to drop. However, the sender and receiver view of a particular connection cannot tell which gateways on the network are congested, and cannot distinguish between propagation delay and queuing delay. Only the gateway has a true view of the state of the queue, the link share of the connections passing through it at any given time, and the quality of service requirements of the
traffic flows. The RED gateway monitors the average queue length, which detects early signs of impending congestion (average queue length exceeding a predetermined threshold) and reacts appropriately in one of two ways:
− Drop incoming packets with a certain probability, to indirectly inform the source of congestion, the source needs to reduce the transmission rate to keep the queue from filling up, maintaining the ability to absorb incoming traffic spikes.
− Mark “congestion” with a certain probability in the ECN field in the header of TCP packets to notify the source (the receiving entity will copy this bit into the acknowledgement packet).
Figure 3. 1 RED algorithm
The main goal of RED is to avoid congestion by keeping the average queue size within a sufficiently small and stable region, which also means keeping the queuing delay sufficiently small and stable. Achieving this goal also helps: avoid global synchronization, not resist bursty traffic flows (i.e. flows with low average throughput but high volatility), and maintain an upper bound on the average queue size even in the absence of cooperation from transport layer protocols.
To achieve the above goals, RED gateways must do the following:
− The first is to detect congestion early and react appropriately to keep the average queue size small enough to keep the network operating in the low latency, high throughput region, while still allowing the queue size to fluctuate within a certain range to absorb short-term fluctuations. As discussed above, the gateway is the most appropriate place to detect congestion and is also the most appropriate place to decide which specific connection to report congestion to.
− The second thing is to notify the source of congestion. This is done by marking and notifying the source to reduce traffic. Normally the RED gateway will randomly drop packets. However, if congestion
If congestion is detected before the queue is full, it should be combined with packet marking to signal congestion. The RED gateway has two options: drop or mark; where marking is done by marking the ECN field of the packet with a certain probability, to signal the source to reduce the traffic entering the network.
− An important goal that RED gateways need to achieve is to avoid global synchronization and not to resist traffic flows that have a sudden characteristic. Global synchronization occurs when all connections simultaneously reduce their transmission window size, leading to a severe drop in throughput at the same time. On the other hand, Drop Tail or Random Drop strategies are very sensitive to sudden flows; that is, the gateway queue will often overflow when packets from these flows arrive. To avoid these two phenomena, gateways can use special algorithms to detect congestion and decide which connections will be notified of congestion at the gateway. The RED gateway randomly selects incoming packets to mark; with this method, the probability of marking a packet from a particular connection is proportional to the connection's shared bandwidth at the gateway.
− Another goal is to control the average queue size even without cooperation from the source entities. This can be done by dropping packets when the average size exceeds an upper threshold (instead of marking it). This approach is necessary in cases where most connections have transmission times that are less than the round-trip time, or where the source entities are not able to reduce traffic in response to marking or dropping packets (such as UDP flows).
3.2.2 Algorithm
This section describes the algorithm for RED gateways. RED gateways calculate the average queue size using a low-pass filter. This average queue size is compared with two thresholds: minth and maxth. When the average queue size is less than the lower threshold, no incoming packets are marked or dropped; when the average queue size is greater than the upper threshold, all incoming packets are dropped. When the average queue size is between minth and maxth, each incoming packet is marked or dropped with a probability pa, where pa is a function of the average queue size avg; the probability of marking or dropping a packet for a particular connection is proportional to the bandwidth share of that connection at the gateway. The general algorithm for a RED gateway is described as follows: [5]
For each packet arrival
Caculate the average queue size avg If minth ≤ avg < maxth
div.maincontent .s1 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 15pt; }
div.maincontent .s2 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: bold; text-decoration: none; font-size: 15pt; }
div.maincontent .p { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; margin:0pt; }
div.maincontent p { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; margin:0pt; }
div.maincontent .s3 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: bold; text-decoration: none; font-size: 14pt; }
div.maincontent .s4 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; }
div.maincontent .s5 { color: black; font-family:"Times New Roman", serif; font-style: italic; font-weight: normal; text-decoration: none; font-size: 14pt; }
div.maincontent .s6 { color: black; font-family:"Times New Roman", serif; font-style: italic; font-weight: bold; text-decoration: none; font-size: 14pt; }
div.maincontent .s7 { color: black; font-family:Wingdings; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; }
div.maincontent .s8 { color: black; font-family:Arial, sans-serif; font-style: italic; font-weight: bold; text-decoration: none; font-size: 15pt; }
div.maincontent .s9 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: bold; text-decoration: none; font-size: 14pt; }
div.maincontent .s10 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 9pt; vertical-align: 6pt; }
div.maincontent .s11 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 13pt; }
div.maincontent .s12 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 10pt; }
div.maincontent .s13 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-d -
Basic Parameters of Single Trawl Fishing Gear -
Basic Techniques (14 Sessions) Attack Techniques (2 Sessions) -
Teaching and learning the title text and epitaph in the 10th grade literature textbook according to genre characteristics - 13 -
Solutions for tourism development in Tien Lang - 10
zt2i3t4l5ee
zt2a3gstourism, tourism development
zt2a3ge
zc2o3n4t5e6n7ts
- District People's Committees and authorities of communes with tourist attractions should support, promote, and provide necessary information to people, helping them improve their knowledge about tourism. Raise tourism awareness for local people.
*
* *
Due to limited knowledge and research time, the thesis inevitably has shortcomings. Therefore, I look forward to receiving guidance from teachers, experts as well as your comments to make the thesis more complete.
Chapter III Conclusion
Through the issues presented in Chapter II, we can come to some conclusions:
Based on the strengths of available tourism resources, the types of tourism in Tien Lang that need to be promoted in the coming time are sightseeing and resort tourism, discovery tourism, weekend tourism. To improve the quality and diversify tourism products, Tien Lang district needs to combine with local cultural tourism resources, at the same time combine with surrounding areas, build rich tourism products. The strengths of Tien Lang tourism are eco-tourism and cultural tourism, so developing Tien Lang tourism must always go hand in hand with restoring and preserving types of cultural tourism resources. Some necessary measures to support and improve the efficiency of exploiting tourism resources in Tien Lang are: strengthening the construction of technical facilities and labor force serving tourism, actively promoting and advertising tourism, and expanding forms of capital mobilization for tourism development.
CONCLUDE
I Conclusion
1. Based on the results achieved within the framework of the thesis's needs, some basic conclusions can be drawn as follows:
Tien Lang is a locality with great potential for tourism development. The relatively abundant cultural tourism resources and ecological tourism resources have great appeal to tourists. Based on this potential, Tien Lang can build a unique tourism industry that is competitive enough with other localities within Hai Phong city and neighboring areas.
In recent years, the exploitation of the advantages of resources to develop tourism and build tourist routes in Tien Lang has not been commensurate with the available potential. In terms of quantity, many resource objects have not been brought into the purpose of tourism development. In terms of time, the regular service time has not been extended to attract more visitors. Infrastructure and technical facilities are still weak. The labor force is still thin and weak in terms of expertise. Tourism programs and routes have not been organized properly, the exploitation content is still monotonous, so it has not attracted many visitors. Although resources have not been mobilized much for tourism development, they are facing the risk of destruction and degradation.
2. Based on the results of investigation, analysis, synthesis, evaluation and selective absorption of research results of related topics, the thesis has proposed a number of necessary solutions to improve the efficiency of exploiting tourism resources in Tien Lang such as: promoting the restoration and conservation of tourism resources, focusing on investment and key exploitation of ecotourism resources, strengthening the construction of infrastructure and tourism workforce. Expanding forms of capital mobilization. In addition, the thesis has built a number of tourist routes of Hai Phong in which Tien Lang tourism resources play an important role.
Exploiting Tien Lang tourism resources for tourism development is currently facing many difficulties. The above measures, if applied synchronously, will likely bring new prospects for the local tourism industry, contributing to making Tien Lang tourism an important economic sector in the district's economic structure.
REFERENCES
1. Nhuan Ha, Trinh Minh Hien, Tran Phuong, Hai Phong - Historical and cultural relics, Hai Phong Publishing House, 1993
2. Hai Phong City History Council, Hai Phong Gazetteer, Hai Phong Publishing House, 1990.
3. Hai Phong City History Council, History of Tien Lang District Party Committee, Hai Phong Publishing House, 1990.
4. Hai Phong City History Council, University of Social Sciences and Humanities, VNU, Hai Phong Place Names Encyclopedia, Hai Phong Publishing House. 2001.
5. Law on Cultural Heritage and documents guiding its implementation, National Political Publishing House, Hanoi, 2003.
6. Tran Duc Thanh, Lecture on Tourism Geography, Faculty of Tourism, University of Social Sciences and Humanities, VNU, 2006
7. Hai Phong Center for Social Sciences and Humanities, Some typical cultural heritages of Hai Phong, Hai Phong Publishing House, 2001
8. Nguyen Ngoc Thao (editor-in-chief, Tourism Geography, Hai Phong Publishing House, two volumes (2001-2002)
9. Nguyen Minh Tue and group of authors, Hai Phong Tourism Geography, Ho Chi Minh City Publishing House, 1997.
10. Nguyen Thanh Son, Hai Phong Tourism Territory Organization, Associate Doctoral Thesis in Geological Geography, Hanoi, 1996.
11. Decision No. 2033/QD – UB on detailed planning of Tien Lang town, Hai Phong city until 2020.
12. Department of Culture, Information, Hai Phong Museum, Hai Phong relics
- National ranked scenic spot, Hai Phong Publishing House, 2005. 13. Tien Lang District People's Committee, Economic Development Planning -
Culture - Society of Tien Lang district to 2010.
14.Website www.HaiPhong.gov.vn
APPENDIX 1
List of national ranked monuments
STT
Name of the monument
Number, year of decisiondetermine
Location
1
Gam Temple
938 VH/QĐ04/08/1992
Cam Khe Village- Toan Thang commune
2
Doc Hau Temple
9381 VH/QĐ04/08/1992
Doc Hau Village –Toan Thang commune
3
Cuu Doi Communal House
3207 VH/QĐDecember 30, 1991
Zone II of townTien Lang
4
Ha Dai Temple
938 VH/QĐ04/08/1992
Ha Dai Village –Tien Thanh commune
APPENDIX II
STT
Name of the monument
Number, year of decision
Location
1
Phu Ke Pagoda Temple
178/QD-UBJanuary 28, 2005
Zone 1 - townTien Lang
2
Trung Lang Temple
178/QD-UBJanuary 28, 2005
Zone 4 – townTien Lang
3
Bao Khanh Pagoda
1900/QD-UBAugust 24, 2006
Nam Tu Village -Kien Thiet commune
4
Bach Da Pagoda
1792/QD-UB11/11/2002
Hung Thang Commune
5
Ngoc Dong Temple
177/QD-UBNovember 27, 2005
Tien Thanh Commune
6
Tomb of Minister TSNhu Van Lan
2848/QD-UBSeptember 19, 2003
Nam Tu Village -Kien Thiet commune
7
Canh Son Stone Temple
2160/QD-UBSeptember 19, 2003
Van Doi Commune –Doan Lap
8
Meiji Temple
2259/QD-UBSeptember 19, 2002
Toan Thang Commune
9
Tien Doi Noi Temple
477/QD-UBSeptember 19, 2005
Doan Lap Commune
10
Tu Doi Temple
177/QD-UBJanuary 28, 2005
Doan Lap Commune
11
Duyen Lao Temple
177/QD-UBJanuary 28, 2005
Tien Minh Commune
12
Dinh Xuan Uc Pagoda
177/QD-UBJanuary 28, 2005
Bac Hung Commune
13
Chu Khe Pagoda
177/QD-UBJanuary 28, 2005
Hung Thang Commune
14
Dong Dinh
2848/QD-UBNovember 21, 2002
Vinh Quang Commune
15
President's Memorial HouseTon Duc Thang
177/QD-UBJanuary 28, 2005
NT Quy Cao
Ha Dai Temple
Ben Vua Temple
Tien Lang hot spring
div.maincontent .p { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; margin:0pt; } div.maincontent p { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; margin:0pt; } div.maincontent .s1 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; font-size: 16pt; } div.maincontent .s2 { color: black; font-family:"Times New Roman", serif; font-style: italic; font-weight: bold; text-decoration: none; font-size: 14pt; } div.maincontent .s3 { color: black; font-family:"Times New Roman", serif; font-style: italic; font-weight: normal; text-decoration: none; font-size: 14pt; } div.maincontent .s4 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; font-size: 14pt; } div.maincontent .s5 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: bold; font-size: 14pt; } div.maincontent .s6 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 14pt; } div.maincontent .s7 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: bold; text-decoration: none; font-size: 14pt; } div.maincontent .s8 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; text-decoration: none; font-size: 9pt; vertical-align: 6pt; } div.maincontent .s9 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: bold; text-decoration: none; font-size: 12pt; } div.maincontent .s11 { color: black; font-family:"Times New Roman", serif; font-style: normal; font-weight: normal; tex
Table 4.1. Experimental results of the models on the CNN/Daily Mail dataset. The symbol '(*)' is the model of See et al. [43]
When the experiment in [43] was repeated using the first 400 words of the article as input, the R-1 score was obtained as 35.87%. However, when using the entire article as input, the R-1 measure dropped to 29.71%. This is because when feeding a long text to the model, the first part of the text is “forgotten” by the system, and the main content of the article is often located at the beginning. However, articles summarized in this way will reduce the generality of the system as well as in cases where important information may not be located in the first 400 words of the text.
Table 4.1 shows that when using the full text of the article as input, both proposed models (model 3 and model 4) outperform the systems in [128] and [43] in all three measures R-1, R-2, and RL. The experimental results show that the sentence position feature is important information in generating a quality summary and word frequency is a good indicator for text summarization tasks using deep learning techniques. When information about sentence position and word frequency are added to the model, the R-1 measure is significantly improved, 2.18% higher than the R-1 measure of the system in [43].
Table 4.2 below shows the experimental results on the Baomoi dataset.
Model
News | |||
R-1 | R-2 | RL | |
Model 1 (Seq2seq + attn baseline) [128] | 26.68 | 9.34 | 16.49 |
Model 2 (Pointer-Generator + Coverage) [43] (*) | 28.34 | 11.06 | 18.55 |
Model 3 ( (*) + POSI) | 29.47 | 11.31 | 18.85 |
Model 4 ( (*) + POSI + TF) | 30.59 | 11.53 | 19.45 |
Table 4.2. Test results of the models on the Baomoi dataset. The symbol '(*)' is the model of See et al. [43]
The results in Table 4.2 also show that both proposed models achieve higher R-1, R-2, RL scores than the other two systems. The best proposed model obtains an R-1 measure that is 2.25% higher than the R-1 measure of the model in [43] and 3.91% higher than the R-1 measure of the baseline model in [128].