American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

30 | P a g e  
 

 

ENERGY EFFICIENT TECHNIQUE AND ALGORITHM 

BASED ON ARTIFICIAL INTELLIGENCE IN CONTENT 

DELIVERY NETWORKS  
Bakhtiyor Makhkamov,  

Nazirjon Khasanov 

Tashkent University of Information Technologies  

Named after Muhammad al-Khwarizmi 

 

Abstract:  

Due to the large number of servers and network infrastructure required to deliver content 

to users, content delivery networks (CDN) consume a large amount of energy. CDN use 

several strategies to reduce energy consumption, such as server consolidation, dynamic 

provisioning, and load balancing. However, these strategies do not take into account the 

popularity of the content being presented. Therefore, a mechanism to improve energy 

efficiency based on content popularity has been developed in CDN. The main function of 

the mechanism is to make maximum use of the cache servers' memory capacity at the 

expense of optimal service to user requests and to increase the service performance of 

cache servers to user requests. To achieve this, based on machine learning algorithms 

using user requests and attributes of video files, predicting the probability of video content 

becoming popular and storing videos with the highest popularity index on edge cache 

servers is caught. 

 

Keywords: Content delivery network (CDN), cache, content, Artificial intelligence, 

Machine learning, predicting, optimizing. 
 

 

Introduction 

Cache Placement and Cache Offloading processes are important in CDN to ensure the speed 

and reliability of content delivery to users around the world. Cache deployment is the 

distribution of CDN cache across geographic regions to increase content availability and 

reach for users. Cache swapping is the process of transferring cached data from a central 

CDN server to servers closer to end users. These are detailed below. 

Caching technologies are used in content delivery networks to distribute video services in 

network segments that are maximally close to users. Since the capacity of caches is limited 

compared to the capacity of servers, they need to store the most requested content. Caching 

highly requested content significantly reduces access latencies, server overload, the amount 

of data to be cached, and resource costs required for caching [1]. 

Figure 1 shows the steps involved in delivering content by caching video services. In the 

typical case of video services, a user makes a video request to a server, and if the requested 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

31 | P a g e  
 

 

video is available in the cache serving the user, the user retrieves the video from that cache. 

If the video is not available in the cache, then the user's video request is forwarded to the 

remote server. Here, the first case is called caching, and the second case is called non-

caching. The cache server updates its caches by replacing existing cached videos with newly 

requested videos, which is controlled by the cache update policy. 

 
Figure 1. Content delivery process 

 

The content delivery process described above has a simple structure and is appropriately 

designed for simple cache servers. In the event that the content of the provided service is not 

available in the local cache server, the user's request is directed directly to the main server, 

and in the process of making such a request, the resources of the network and other related 

elements costs. In the case of caching with a complex structure, the process of caching the 

content of the provided service is carried out hierarchically, which maximizes the caching 

priorities. 

Caching has been widely used in content-based networks. Content placement algorithms are 

used to determine the order of routers from the actual content source to the user. 

Operational cache placement algorithms are important for Internet Service Providers (ISP) 

to reduce traffic and achieve good CDN performance. LCE (Leave Copy Everywhere), LCD 

(Leave Copy Down) and Prob algorithms are the last caches to determine the best in terms 

of lead time. 

 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

32 | P a g e  
 

 

Materials and methods 

Energy efficient VoD in content delivery network. Caching content locally results in 

a shorter route to content and lower power consumption as a result. However, this strategy 

leads to increased device power consumption by deploying local caches. Therefore, the 

optimal cache size is a function of the above two metrics, and the optimal value between 

them should be found. This evaluation aims to minimize the power consumption of the video 

service by optimizing the size of the caches deployed in the nodes. This considers a Video-

on-Demand (VoD) service combining IP over a WDM network with the network architecture 

illustrated in Figure 2. 

 
Figure 2. WDM-based content delivery architecture 

VoD is an energy-intensive service due to the high energy consumption associated with the 

storage and delivery of large multimedia files. Keeping the most popular content close to the 

network boundary is one effective way to reduce the energy consumption of video services. 

We have estimated the energy consumption of video over IP in a WDM network. It explores 

the power savings introduced by using caches in the core nodes. Using a linear programming 

model, it is designed to optimize the size of the cache at each node in the network in order 

to minimize power consumption. The linear programming model is extended to account for 

variable caches, meaning that caches equipped with sleep capabilities allow inactive parts to 

be turned off when traffic is low. The model finds the optimal cache size for each node for 

different times of the day, which allows to achieve the best energy efficiency [2-3]. 

A variable-cache linear programming model finds the optimal cache size of each node that 

varies throughout the day. In the model, caches are assumed to have sleep capabilities, so 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

33 | P a g e  
 

 

inactive portions of the cache can be put to sleep. The goal is to explore potential additional 

power savings while using fixed caches and analyze changes in the optimal cache size that 

minimizes power consumption as network traffic changes. 

This linear programming model finds the optimal cache size to be placed at each node in the 

network to minimize power consumption. Caches are assumed to have full working memory 

throughout the day. 

The router ports required for each band wavelength at time t are determined by expression 

(1): 

 
  
















+
Ni ji

ijtitp

Nmj
i

CAPP
:

        (1) 

where: 𝐴𝑃𝑖𝑡 – aggregation ports on node i at time t; 𝐶𝑥𝑦𝑡 – the wavelengths in the virtual 

connection from node x to node y at time t. 

The power consumption of the optical switch at time t is determined by expression (2): 

it
i N

Po


       (2) 

where: 𝑃𝑜𝑖𝑡 – the power consumption of optical switch i at time t. 

transponders at time t: 

i

ijt
i N i Nm

Pt w
 
       (3) 

amplifiers at time t: 

i

ijt ijt
i N i Nm

Pa Amp f
 
       (4) 

Multiplexers/demultiplexers at time t: 

i

ij
i N

Pmd

i Nm
f

 
       (5) 

Caches installed at time t: 

i N

M


      (6) 

It should be noted that the model does not assume a simple symmetric case, and therefore 

the number of lightpaths from node i to node j may differ from the number of lightpaths in 

the opposite direction. Basically, fij, wijt, and Cijt are not necessarily equal to fji, wjit, and Cjit, 

respectively. In this case, uplink traffic is video traffic uploaded from nodes to video servers, 

downlink traffic is video traffic downloaded from video servers to nodes, and regular traffic 

is traffic that cannot be cached (e-mail, live video, dynamic content, etc.) 

The objective of the proposed integrated linear programming model is to minimize the daily 

power consumption of the network while overcoming a number of current and power 

constraints. A complete linear programming model is defined as: 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

34 | P a g e  
 

 

 :
i i

i i

p it ijt it ijt
i N i N i Ni j

t T

itij ij ij
i N i N i N

Pt

j Nm j Nm

Pa Pmd

j Nm j Nm

C Po WP AP

Amp f f M

  



  

  
  + + + +
     
 + + 
 

  

    


    

  (7) 

Through expression (7), the power consumption of the network can be calculated by 

summing the power consumption of different network components at each time point. The 

calculation of the optimal value of this analytical model is carried out based on the 

mechanism presented in Figure 3. 

 
Figure 3. Energy efficiency technique based on content popularity in CDN 

 

As can be seen from the expression (7), all indicators except M are constant parameters and 

depend on the parameters of the data transmission network. For the most part, these 

parameters are fixed and almost impossible to change. The solution of the problem from this 

point of view is related to the problem of optimization of Mit - the amount of caches installed 

at time t. From this point of view, the main result of the mechanism presented in Figure 3 is 

to maximize the use of cache servers' memory capacity at the expense of optimal service to 

user requests and to increase the performance of cache servers to user requests. To achieve 

this, based on machine learning algorithms using user requests and attributes of video files, 

predicting the probability of video content becoming popular and storing videos with the 

highest popularity index on edge cache servers is caught [4-6]. 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

35 | P a g e  
 

 

A network consists of a central server, N:={1,...,N} cache servers, and a number of users 

associated with each cache. To simplify the process, we consider the case where each user is 

connected to a single cache server. Although only one central server is considered, the 

existing formula can be easily extended to a network infrastructure with multiple central 

servers. File requests from users are received on cache servers and served locally or 

forwarded to a central server depending on file availability. Each cache server stores only a 

portion of the entire content library. The purpose of this is to define the set of files that 

should be stored on the cache servers. 

Below is a general CDN setup consisting of a central server and N cache servers. i – storage 

size of the cache server is denoted by Mi and is measured in bits. The central server has a 

library of files indexed as 𝑓𝜖𝐹: = {1, . . . 𝐹}. The size of an f file is defined in sf, which is also 

measured in bits. File sizes are collected into a 𝑠𝜖𝑅𝐹++ vector. The time is divided into slots 

and the time slot is indexed as t=1,2,.... The set of files stored in the i-cache at time t is 

denoted by 𝐹𝑖(𝑡)𝜖𝐹. Alternatively, the storage vector 𝑚𝑖(𝑡) for the i-cache is defined as 

𝑚𝑖𝑓(𝑡) = 1 for 𝑓𝜖𝐹𝑖(𝑡), otherwise zero. If the storage vector mi(t) satisfies 𝑠𝑇𝑚𝑖(𝑡) ≤ 𝑀𝑖, it is 

considered valid for i-cache. Let Ui be the set of users who send requests to i-cache. Since 

different servers may have limited network capacity and service ratio, it is assumed that 

users in Ui can receive Cc i bits and Cr i bits from the central server in each time slot from 

the i-cache at the maximum total rate in each time slot. The total cache request i for file f in 

each time interval t is given by dif(t), which is measured in bits and summed into the vector 

𝑑𝑖(𝑡)𝜖𝑅𝐹 +. Given a given demand dif(t), if mif(t)=1, the service policy performs service 

locally and redirects the request otherwise. Assuming that all requirements are subsequently 

satisfied, the total flow rate from the i-cache server is given by 𝑚𝑇𝑖(𝑡)𝑑𝑖(𝑡), and the flow rate 

from the central server to users at Ui is given by the expression 1𝑇 𝑑𝑖(𝑡) − 𝑚𝑇𝑖(𝑡)𝑑𝑖(𝑡) is 

given, measured in bits in both timeslots. 

If user requirements are known in advance, optimal content placement involves solving an 

integer programming problem in mi(t). The idea is to associate relevant costs with network 

flows and find the file placement that minimizes the overall cost. In the present context, we 

define the cost functions corresponding to the flows in the last mile connections connected 

to the i-cache server as χi:R→R and corresponding to the flows between the user connected 

to the i-cache server and the central server φi:R→R must be determined (blue dotted lines 

in Figure 3). On the other hand, for the topology presented in Figure 3, φi(.) represents the 

cumulative cost function that penalizes the flows in the backlinks and last-mile connections 

connected to the i-cache. Since the cost functions can usually be arbitrary, this approach is 

significantly more general and provides more flexibility than traditional displacement-based 

policies. As an example, the quadratic penalty 𝜒𝑖(𝑥) = 𝑐1𝑥2 + 𝑐2 imposes a base cost in 

addition to the square of the network usage, which prevents very high usage on any single 

server, thereby preventing buffer overflows or requests prevents the possibility of deletion. 

Another example is a simple linear cost that is multiplied by 𝜒𝑖(𝑥) = 𝑤𝑥 + 1 2x2, with a strict 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

36 | P a g e  
 

 

convex regularization in addition to penalizing large file downloads. Finally, the Kleinrock 

average delay function for a narrow link with capacity Ci is given by the expression 𝜒𝑖(𝑥) =

𝑥/(𝐶𝑖 − 𝑥) for 0<x<Ci. In contrast, eviction-based policy parameters typically cannot directly 

control delays [7-8]. 

Flexibility in response to temporary changes in demand is also important in practice. For 

example, the requirements {𝑑𝑖𝑓(𝑡)} ≥ 1 are usually independent and not uniformly 

distributed, but exhibit temporal decay and correlation. In particular, the popularity of 

content often decreases over time. For example, it has been observed that demands often 

have a diurnal approach and reach peak values at certain times of the day. A complete 

implementation of the content caching algorithm is shown in Figure 4. 

 
Figure 4. Content caching algorithm 

 

In this case, the requirements are accumulated over time 𝑇𝜆 and used to fill the matrix D(t). 

At each update interval, the expected stream variables {𝑥′𝑖𝑓 , 𝑦′𝑖𝑓}  are obtained using the 

current value of 𝜆𝑖𝑓 in the given algorithm, which is then updated. Expected streams are also 

used to update storage vectors {𝑚𝑖(𝑡)}. The function cache(𝑋(𝑡), 𝑌(𝑡))  indicates whether one 

of the placement policies is used. The cache refresh interval is determined by a separate size. 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

37 | P a g e  
 

 

To get an additional boost on the performance of the algorithm, it is convenient to think of 

λ as the hidden cost of not satisfying the demand. Prices increase when demand is not 

satisfied by expected flows at any time t or when 𝑥𝑖𝑓(𝑡) + 𝑦𝑖𝑓(𝑡) − 𝑑𝑖𝑓(𝑡) is negative. 

Expected flows at key upgrade stages are adjusted based on higher or lower shadow prices. 

For example, if the demands are constant, the expected flows follow them so that the 

difference 𝑥𝑖𝑓(𝑡) + 𝑦𝑖𝑓(𝑡) − 𝑑𝑖𝑓(𝑡) is not very large. 

As a simple example, consider a file f where the demand i from the user is zero for all time t. 

For such a file, if 𝑙𝑖𝑓(0) = 0, it remains zero for all time t. In fact, even though the 

initialization is different, the dual variables approach zero because 𝑥′𝑖𝑓(𝑡) and 𝑦′𝑖𝑓(𝑡) are 

fixed whenever 𝑙𝑖𝑓(0) > 0 will be positive. It can also be said that if a given dif(t) remains 

high for all t. Extending the argument, the double variable 𝑙𝑖𝑓(𝑡) “observes” the requirements 

of 𝑑𝑖𝑓(𝑡), although its observability is limited by the step size µ. Thus, the expected streams 

{𝑥′𝑖𝑓(𝑡), 𝑦′𝑖𝑓(𝑡)}, which are monotone functions of 𝑙𝑖𝑓(𝑡), also track the popularity of files. In 

addition, occasional increases or decreases in demand have a minimal effect on the evolution 

of 𝑙𝑖𝑓(𝑡) due to the averaging effect inherent in the dual algorithm [8-9]. 

The interpretation presented here forms the basis of the proposed content placement 

algorithm. Expected streams represent the average popularity of files that can be used to 

host content online. 

 

Results and Discussion 

When serving users on a traditional CDN, after receiving a user request, the cache server 

checks whether the file exists as a local element. In case of a cache miss, the file is 

downloaded from the central server, served to the user, and stored on the cache server. 

When the cache is full, different migration algorithms are used, which are explained in 

detail. 

In this work, in contrast to schemes based on permutation, a method using expected flows 

– {𝑥𝑖𝑓(𝑡)} is used to obtain the memory allocation mi(t). Because pending streams are 

designed to meet average, not instantaneous, requirements, the resulting memory 

allocations typically do not contain every file that is eligible for a cache miss depending on 

the value of 𝑥𝑖𝑓(𝑡). This approach differs greatly from existing replacement algorithms such 

as LRU, where the out-of-cache file is always saved and the saved file is deleted. 

Deployment of content within a shared first topology should be done periodically because of 

the high cost of retrieving the associated content. The refresh interval marked as tv up varies 

from hours to days, and files in the cache are updated at the beginning of each such interval. 

For this, another content placement algorithm has been developed, its implementation is 

shown in Figure 5. 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

38 | P a g e  
 

 

 
Figure 5. Content optimal caching algorithm based on artificial intelligence 

 

Developed algorithm: First, files with high 𝑥𝑖𝑓(𝑡) values are considered popular and cached. 

That is, with each update, the cache ensures that the files with the largest 𝑥𝑖𝑓(𝑡) values 

match. In this case, the popularity of files is predicted based on artificial intelligence 

methods. As file sizes may vary, the number of saved files may also vary. If the list of files 

requested in a recent time interval is very large, only the files matching the largest values of 

𝑥𝑖𝑓(𝑡) are placed in the cache edges. 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

39 | P a g e  
 

 

In the developed algorithm, the cache server provides video content based on the requests 

received by the user. In this case, if the requested content is available in the cache (1) the 

video is presented from the cache, otherwise, the video is presented in the form of streaming 

from the central server through the cache (2). In both cases, the expected values of 𝑥𝑖𝑓(𝑡) of 

the streams of files in the cache are determined. At the same time, based on video attributes 

on the cache server and internal file attributes, content popularity prediction algorithms 

work on the central servers (3) and calculate the probable popularity coefficients Km of the 

existing files [9-10]. 

The cache servers operate without updating the stock of files for a pre-entered Tv up time, 

and at the specified time it sends a request to download the videos with the highest 

probability of being popular (4). Based on the size of the cache, the files whose popularity is 

higher than the value 𝑥𝑖𝑓(𝑡) of the files in the cache are downloaded, and the cache files are 

replaced based on the condition Km>𝑥𝑖𝑓(𝑡) (5). This algorithm ensures that video content 

that can be requested by users many times is placed on the cache server and that future video 

requests are served based on the maximally limited caches. As a result, the volume of 

information exchange between users and central servers is minimized. This, in turn, allows 

you to increase the Hit ratio and, in turn, make maximum use of the available cache memory. 

Through this, the number of requests to the central server or external servers is reduced and 

the overall network energy saving is increased. 

 

Energy consumption of CDN network elements 

The main elements of a CDN network are a central server, routers, switches, transponders, 

amplifiers, multiplexers/demultiplexers, and edge cache servers. Therefore, in order to 

improve the energy efficiency of the CDN network, primary information about their energy 

consumption is needed. Below are the realistic energy consumption values of these 

elements. 

Central server. Usually, server equipment manufacturers provide information about energy 

consumption in technical documents or specifications. Power consumption can be set for 

different operating modes, including maximum load (peak demand), normal load, or 

standby load. 

Routers and their ports. The energy consumption of routers depends on several factors, such 

as: model/manufacturer, configuration, intensity of use, data transmission technologies. In 

general, the power consumption of a router ranges from a few watts (W) to several hundred 

watts, based on the above facts. As an additional note, the total power consumption of a 

router also depends on the number of ports required per wavelength. The number of ports 

varies depending on the wavelength division technology (TDM, WDM) used. Thus, the 

actual number of router ports required for each band wavelength depends on the router's 

capabilities. Watt is expressed as (8): 

1W=B*A      (8) 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

40 | P a g e  
 

 

Based on the above information, the energy consumption of routers is determined from 5W 

to 12W per hour. 

Switches and their ports. The energy consumption of switches is almost the same as the 

energy consumption of the router mentioned above. Therefore, their energy consumption is 

determined from 5W to 12W per hour. 

Transponders. The energy consumption of transponders in the CDN network can vary 

significantly depending on the technologies used, models and the amount of data 

transmitted. On average, the power consumption of a transponder usually varies from 

several tens to several hundred watts. However, it should be noted that more accurate 

figures can only be provided by the manufacturer or seller, since the exact power 

consumption depends on the specific transponder model and configuration, as well as the 

current workload and operating conditions. For more accurate information on power 

consumption, it is recommended to refer to the specifications or technical documentation of 

a particular transponder. The energy consumption of common optical transponders does 

not exceed 200 watts per hour. 

Amplifiers. The power consumption of amplifiers in a CDN network can also vary 

significantly depending on the specific conditions and characteristics of the device. Usually, 

the power consumption of the amplifier is from several tens to several hundred watts. 

However, exact values may vary greatly depending on factors such as amplifier type and 

model, amplification technologies used, device configuration and operating mode. For 

example, optical amplifiers used in fiber optic communication systems can have a power 

consumption of 10 to 100 watts per hour. 

Multiplexer/demultiplexers. Power consumption of multiplexers and demultiplexers is 

measured in watts. This depends on various factors, including the specific model, technology 

and operating frequency of the device, as well as the number of I/O and data being 

converted. In general, the power consumption of a multiplexer/demultiplexer is 15 to 40 

watts per hour. 

Edge cache servers. The energy consumption of a marginal cache server is 100 to 250 watts 

per hour for a medium-sized company. Such a server can have two or four processors, the 

amount of RAM varies from 16 to 32 GB, and several hard drives from 1 to 4 TB each. A 

linear programming model can be implemented using the energy consumption data of the 

CDN network elements presented above. 

The input values for the power consumption parameters used in the model are shown in 

Table 1. The power consumption of an 8-slot CRS-1 is 8000 W, including backplane power 

consumption, and therefore the power consumption of a 40 Gb/s port is estimated to be 

1000 W. The power consumption of the Cisco ONS 15454 optical filter card is 16 W and the 

Cisco ONS 15501 optical amplifier is 8 W. 

 

 

 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

41 | P a g e  
 

 

Table 1. Input values for power consumption parameters 

The distance between two adjacent EDFAs (S) 80 (km) 

The number of wavelengths in a fiber (W) 16 

Wavelength capacity (B) 40 (Gbps) 

Router port power consumption (Pp) 1000 (W) 

Transponder power consumption (Pt) 73 (W) 

EDFA power consumption (Pa) 8 (W) 

Optical switch power consumption (Po) 85 (W) 

MUX/DEMUX power consumption (Pmd) 16 (W) 

Cache server power consumption factor (∅) 7,4 (WpGB) 

Probability of hitting the cache server Avoid/fall/proposed method 

 

The energy consumption of a 1-bit stream (READ/WRITE operation) is given as 211×10-9 J. 

This value is converted to the power (watts) used to transfer 1 GB of data over a certain 

period of time for use in the analytical model. Energy consumption for streaming 1 GB of 

data: 211×10-9 ×109 ×8=1688 J. A typical maximum access speed for a Fiber Channel hard 

drive is up to 100 MBps. However, typical average data rates are about 50% of this value and 

reach up to 50 MBps. Increasing the access speed of the hard disk will reduce the cache 

refresh time, but will result in more power consumption. To achieve average power 

consumption, a data transfer rate of less than 35 Mb/s is considered on the hard disk. Time 

required to stream 1 GB: 103×8/35=228 seconds. So, the power consumption for 1 GB of 

caching is 1688/228 = 7.4 WpGB. 

 
Figure 6. Energy consumption as a percentage of whether user requests are cached or not 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

42 | P a g e  
 

 

The result in Figure 6 is realized by the analytical model, which takes into account the 

variation of the cache index. The change in the cache indicator shows the probability that 

users will not access the content requested by the local cache server, will access it (existing 

methods), and will access it through the proposed method. The result is the energy 

consumption due to increasing the size of the local cache server, determined by whether or 

not the cache is dropped. According to the obtained results, a steady 74 watts of power 

consumption in non-dropping, 38-51 watts of power consumption with existing methods 

with a cache server, and 28-31 watts of power consumption with the proposed method were 

achieved. The efficiency of the proposed method is 40% better than the existing methods. 

 

Conclusion 

A mechanism for efficiently placing content in CDN caches and an algorithm based on 

artificial intelligence have been developed. This mechanism and an algorithm based on 

artificial intelligence aim to improve the efficiency of content placement in CDN caches, 

taking into account several factors such as content popularity, user location and server 

availability. The mechanism consists of the following steps: content popularity analysis; 

analyze user location; analysis of server availability; and applying an algorithm based on 

artificial intelligence. In content delivery networks, the number of requests to central or 

external servers and the total energy consumption of the network have been reduced 

through the mechanism and algorithm of effective caching of video content based on 

artificial intelligence. Analysis of the energy consumption of the elements of the CDN 

network was carried out, based on the results of the analysis, calculations of the energy 

consumption were carried out, taking into account whether the user's requests were 

delivered to the cache server or not. The result is the energy consumption due to increasing 

the size of the local cache server, determined by whether or not the cache is dropped. 

According to the obtained results, a steady 74 watts of power consumption in non-dropping, 

38-51 watts of power consumption with existing methods with a cache server, and 28-31 

watts of power consumption with the proposed method were achieved. The efficiency of the 

proposed method is found to be 40% higher than the existing methods. 

 

References 

1. V. Mathew, R. K. Sitaraman, and P. Shenoy, “Energy-aware load balancing in content 

delivery networks,” in 2012 Proceedings IEEE INFOCOM, 2012, pp. 954–962. 

2. W. Wang et al., “CRCache: Exploiting the correlation between content popularity and 

network topology information for ICN caching,” in Communications (ICC), 2014 IEEE 

International Conference on, 2014, pp. 3191–3196. 

3. Liangzhong Yin and Guohong Cao, “Supporting cooperative caching in ad hoc networks,” 

IEEE Trans. Mob. Comput., vol. 5, no. 1, pp. 77–89, 2006. 



 

American Journal of Interdisciplinary Research and Development 

ISSN Online: 2771-8948 

Website: www.ajird.journalspark.org 

Volume 20, Sep., 2023 

43 | P a g e  
 

 

4. G. P. Lorenzo Saino, Ioannis Psaras, “Icarus: a Caching Simulator for Information Centric 

Networking (ICN),” in . ICST (Institute for Computer Sciences, Social-Informatics and 

Telecommunications Engineering)., 2014, pp. 66–75. 

5. A. Compagno, M. Conti, C. Ghali, and G. Tsudik, “To NACK or not to NACK? negative 

acknowledgments in Information-Centric Networking,” in Computer Communication and 

Networks (ICCCN), 2015 24th International Conference on, 2015, pp. 1–10. 

6. W. K. Chai, D. He, I. Psaras, and G. Pavlou, “Cache 'Less for More' in Information-Centric 

Networks,” pp. 27–40, 2012. 

7. Xiaohu Chen, Qilin Fan, and Hao Yin, “Caching in Information-Centric Networking: From 

a content delivery path perspective,” in Innovations in Information Technology (IIT), 2013 

9th International Conference on, 2013, pp. 48–53. 

8. Y. Wu, Y. Jiang, M. Bennis, F. Zheng, X. Gao, and X. You, “Content popularity prediction 

in fog radio access networks: A federated learning based approach,” in 2020 IEEE 

International Conference on Communications (ICC), Jun. 2020, pp. 1–6. 

9. Abdorasoul Ghasemi, Amirhosein Ahmadi, Cache management in content delivery 

networks using the metadata of online social networks, Computer Communications, Volume 

189, 2022, Pages 11-17, ISSN 0140-3664, https://doi.org/10.1016/j.comcom.2022.02.021. 

10. Y. Tao, Y. Jiang, F. Zheng, M. Bennis, and X. You, “Content popularity prediction in fog-

rans: A bayesian learning approach,” in 2021 IEEE Global Communications Conference 

(GLOBECOM), Dec. 2021, pp. 1– 6.