Toward IoT fog computing-enabled system energy consumption modeling and optimization by adaptive TCP/IP protocol

Used applications to realize the experiments

1. FTP, file transfer protocol, RFC 959. FTP was used to transmit a continuous stream of data, to see its effect on the energy consumed in sending data.

2. A Java program was used as an alternative to the FTP protocol. The same idea of sending a continuous stream of data to see its effect on the energy consumed in sending data. As the FTP protocol, client–server programs were used on both machines. Another use of Java was in measuring the silence time, where a do-nothing program was used.

3. Ping, Packet Internet Groper. A ping storm is a condition in which the internet ping program is used to send a flood of packets to a machine to test its ability to handle a high amount of traffic. This is known in ping options as –f option, which was used in this work. Some other options, like –s to define the size of packets or –c to define the count of echo request packets, were also used.

Note: the used system is Linux Ubuntu, and to use the –c option user must be root to define the count to be more than three packets.

Measuring principles

The following steps explain the used techniques to measure the consumed energy:

1. Measuring the silence time that is the time in which the machine does not send or receive any data. To do the experimentation, either the time or the consumed energy must be fixed to a certain value. In this work, the energy consumed was always 25% of the full battery energy. The battery was charged until 100% and the measured values were until the battery reached 75% (25% of energy was consumed).

2. Measuring the consumed energy from the system itself. This was done by deactivating the transmission antenna to ensure that no data is sent or received. To accomplish this point, a do-nothing java program was turned on the machine.

3. The program was left running until the device dropped from 75% to 25% of the battery level. These values were chosen to satisfy the recommendations of the manufacturer. Hence, the minimum recommended charge level is 20% to prevent fast drainage (Hannan et al., 2018). Also, all the employed devices have new batteries (less than six months).

4. To measure the energy consumed in sending and receiving a byte, Ping and ftp protocols were used for a continuous stream of data. In addition, the time spent to reach the required battery level was measured. Here, to calculate the value of the consumed energy for sending or receiving a byte, the energy consumed from the system must be subtracted from the total energy measured for each case. Note that consumed energy is a function of streamed data and time.

5. The measurements were repeated many times, and for different MTU sizes, to get an average value for each device.

A. Measuring the consumed energy

1. Silence Time Consumed Energy: The silence time is the time where the machine does not send or receive any data. The idea here is to measure the time needed for the machine to reach 75% or to consume 25% of the full battery energy. The silence time does not depend on the size of the MTU because no sent or received data or antenna is working. So, for all values of MTU, the silence time will always be the same. Firstly, the battery was fully charged. Then, a do-nothing Java program was turned on the machine. When the battery value reached 75% of the full energy, the time was measured.

Definition: The system energy consumption S, can be expressed in the battery unit per second, as in (1). (1)${\displaystyle S=\frac{25}{d_s}}$ where d_s is the time that the system needs to reach 75% of the full battery for a given MTU size. Table 1 shows the measured time.

The average value of S is: ${\displaystyle S=\frac{25}{d_s}=\frac{25}{3445}=7.26E-3\mathrm{ub}∕\mathrm{s}}$

(2) Transmission Consumed Energy:

The energy consumption in transmitting data is the energy consumed in sending continuous data until reaching 75% of the fully charged battery energy. To measure this value, FTP protocol and a java program were used. A client/server method is used, where one machine was set as a server and another one as a client.

The first step was by opening a connection channel, then a continuous stream of data was sent from the wireless machine to the other machine. The wireless machine was left until the battery reached 75% of the full battery energy. Both the elapsed time and the sent data were measured in each trial. The experiment was held for different values of MTU; 1,500, 1,000, and 500 bytes size. As a result, the energy consumption by the transmission t (m) function for FTP or Java is calculated. Hence: the energy consumed by the system itself must be subtracted from the resulted values.

Table 2 shows the values of d_f(m), dj (m), q_f(m), and q_j(m) for different sizes of MTU. Where d_f(m) and d_j(m) are time in seconds, while q_f(m) and q_j(m) are the quantity of sent data in bytes. Table 3 and Fig. 5 show the variation of energy consumed for different values of MTU.

Definition: The consumption in transmission function t (m) is the energy needed to transmit one byte. It is expressed as a battery unit per byte (octet). From the measured values, the function of t (m) can be expressed as follows: (2)${\displaystyle t\left(m\right)=\frac{25-S.d_x\left(m\right)}{q_x\left(m\right)}}$

From Figure 5, a linear equation for the consumed energy in transmitting one byte can be estimated: (3)${\displaystyle t\left(m\right)=-5.190E-12\ast m+1.464E-08}$ (3) Reception Consumed Energy:

The reception consumed energy is the energy consumed in receiving a continuous stream of data until the battery reaches 75% of its full charge. To realize this experiment, the Ping program was used. In the Ping program, the echo request size equals that of the echo-response. In other words, if a packet of 1,000 bytes was sent as an echo request from one machine, the other machine will reply by sending another packet of 1,000 bytes as an echo response.

Repeatedly, the experiment was to measure the elapsed time and the sent data when the consumed energy is 25%. Table 4 shows the results of this experiment.

Since the consumed energy in sending one byte and the silence time are already known, and it is known that the sent data equals the received data, then this information can be applied with the measured values to obtain the consumption in reception r (m) as follows in Eq. (4): (4)${\displaystyle r\left(m\right)=\frac{25-S.d_p\left(m\right)-q_p\left(m\right).t\left(m\right)}{q_p\left(m\right)}}$

Definition: The consumption in reception function r (m) is the energy needed to receive one byte.

Table 5 and Fig. 6 show the values of r (m). From Fig. 6 a linear estimation equation for r (m) can be obtained: (5)${\displaystyle {r\left(m\right)=-1.806\ast 10}^{-12}{\ast m+1.271\ast 10}^{-8}ub∕o}$

B. Energy consumption estimation and real traffic

In the previous part, the whole work was performed for a continuous data stream with a predefined MTU value for one device. However, in real traffic, the received and sent packets are not always continuous, and their size varies over time.

Nevertheless, the following points have to be considered in real traffic:

1. The Internet is usually not used for a long time.

2. Even when the Internet is used for a long time, the wireless device often receives or sends little data during this time.

3. Most of the received packets are in the form of acknowledgment or a broadcast, and most of them have a small size.

These points summarize the major differences between real traffic and continuous data stream.

In the next part, it is considered that the user would use the Internet not more than 25% of the total usage of his machine. The rest of the usage would be for other things; idle, exchanging control messages, etc. In the case of sensor networks, long periods could separate between occurrences of important data.

Global traffic.

The global traffic g (p) is the number of packets of a given size per second (or unit of time), where p is the packet size. The number of packets of a given size can be zero or more in a given period depending on the traffic itself and the MTU value. In addition, the machine would not send or receive any packet greater than the MTU in size. If the MTU size was 1,500 bytes (which is the default value), there is no packet of size greater than 1,500 bytes. However, if the value of MTU is changed to a new value p, the machine would not receive or send packets greater than p bytes.

In this part, the global traffic was observed by the use of the Linux command tcpdump with option –e. This command allows observing traffic for a given machine, and the –e option prints the link-level header on each dump line. Nevertheless, the experiment can be done only for an MTU size of 1,500 bytes. This is because the MTU size is by default configured to 1,500 bytes for the machines. Thus, it is difficult even impossible to change the size of MTU for all the machines in the network.

To get the real packets in the network, the experiments were held for one hour each, and with the help of tcpdump, grep and sed commands. Figure 7 shows the global traffic (packets received and sent) over the network for one hour and an MTU size of 1,500 bytes. The distribution shows clearly that most of the packets are of small size and in the range 60–160 bytes (nearly 64% of the packets are 60 bytes size). This huge number of small packets would cost the device more energy. This is obvious as shown in Figs. 5 and 6 and with the linear Eqs. (3) and (5). In other words, the consumed energy is proportional to the number of small packets.

By taking a deep look at these packets, it is noticed that the majority of these packets were: Broadcast / arp whohas, acknowledgment, or LLC. These packets could not be neglected or eliminated because they are control packets. Moreover, they will be increased in case the size of MTU is decreased because there will be more received packets and there is a need to send more control packets (especially acknowledgment packets). The next part will focus on this issue.

Transmitted traffic by one machine.

The transmitted traffic by one machine f (p) is the number of packets of a given size per second, (or unit of time), or the packets that are sent by one machine only. The main difference between the global traffic and this traffic is that the size of MTU can be changed for that machine for different values. This helps to observe the transmitted traffic for other MTUs than the 1,500 bytes. In order to observe f(p), tcpdump was used. Once more, the experiments were performed for the three different values of MTU: 1,500, 1,000, 500 bytes. Each experiment was held many times. Figures 8A–8C show the observed transmitted packets on one machine, for one hour.

Again, from the packets’ distribution, it is clear that most of the packets in the range of 62–88 bytes (96% of the transmitted packets). Here it is noticed that the number of 66-byte packets is increased when the MTU size is decreased. This is because more packets have to be sent and received for the same amount of data. Moreover, the acknowledgment packets will increase too (one acknowledgment packet for each received packet). In the next part, the real traffic energy consumption will be discussed.

C. MTU modification over real traffic

Since the value of MTU could not be changed for all the machines on the network, MTU transformation can be done to give an estimation of the energy consumption for different values of MTU. Graphs and packet distribution can help in doing these transformations. However, it is not useful to use the packet distribution as is. To be able to use this distribution, the distribution must be normalized to get a new distribution where the area under the curve equals 1.

The normalization is done by dividing the number of packets of all sizes, to get a normalization function f_i(p), as shown in Eq. (7): (7)${\displaystyle f_i\left(p\right)=\frac{g_i\left(p\right)}{\sum _{j=1}^{m}j\ast g_i\left(j\right)}}$

Where i represents the function g (p) and f (p) for a given MTU, and j is the size of the observed packets.

With the help of the fi (p) functions, a transformation function τ is achieved to estimate the global traffic for a given MTU, and the whole network. To estimate the global traffic for a given value of MTU, it is enough to apply the transformation function on the global function g (p). The idea of this algorithm is to obtain new functions f_n from the function f ₁. Noting that this transformation must preserve the quantity of traffic.

The algorithm that satisfies the previous conditions can be as the following: ${\displaystyle \tau \left(\mathrm{m}2,\mathrm{m}1\right)\left(\mathrm{p}\right)=\left\{\begin{array}{c}0,\mathrm{if}\mathrm{p}>\mathrm{m}2\hfill \\ f\mathit{(}p\mathit{)}+\sum _{i=m1+1}^{m1-1}f\mathit{(}i\mathit{)}\mathit{.}s\mathit{(}i,p,m2\mathit{)},\mathrm{if}\mathrm{p}>\mathrm{m}2\hfill \\ f\mathit{(}p\mathit{)}+\mathrm{int}\left(\frac{f\mathit{(}m1\mathit{)}\mathit{\ast }m1}{m2}\right)+\sum _{i=m2+1}^{m1-1}f\mathit{(}i\mathit{)}\mathit{.}\mathrm{int}\left(\frac{i}{m2}\right),\mathrm{if}\mathrm{p}=\mathrm{m}2\hfill \\ \hfill \\ \text{special}\text{cases}:\hfill \\ \hfill \\ f\mathit{(}p\mathit{)}=f\mathit{(}p\mathit{)}+1,\mathrm{if}\mathrm{p}=\left(\mathrm{f}\left(\mathrm{m}1\right)\mathrm{\ast \; m}1\right)\text{\%}\mathrm{m}2\hfill \\ f\mathit{(}66\mathit{)}=\mathrm{int}\left(\frac{f\mathit{(}66\mathit{)}\mathit{\ast }m1}{m2}\right),\mathrm{to}\text{represent}\mathrm{the}\mathrm{ack}\hfill \end{array}\right.}$ Where: ${\displaystyle \mathit{s}\left(\mathrm{i},\mathrm{p},\mathrm{m}2\right)=\left\{\begin{array}{c}\mathrm{0},\mathrm{if}\mathrm{p}\ne \mathrm{i}\text{\%}\mathrm{m}\mathrm{2}\hfill \\ \mathrm{1},\mathrm{if}\mathrm{p}=\mathrm{i}\text{\%}\mathrm{m}\mathrm{2}\hfill \end{array}\right.}$ m ₁ is the old value of MTU,

m ₂ is the new value of MTU

This algorithm can be summarized in the following points:

1. The size of the new MTU must be less than the size of the old MTU. More information exists to derive a new distribution for smaller MTUs. Nevertheless, the process in the opposite way is more difficult (not done).

2. In the new distribution, the packets that are greater than the size of the new MTU were eliminated. However, they will be divided by the new MTU size. The resulting integer quotient and the remainder are counted as new packets and added to the packets of smaller size

3. The division process is based on multiplying the packets that have a size greater than the new MTU, by its number. The resulting value is divided by the size of the new MTU. The modal of this division is added as one packet to its equivalent packet size.

4. Since it is known that the number of packets will increase, the number of acknowledgments must be also increased. As shown in point B.2, it is the packets of 66-byte size.

Figs. 9A–9C show the generated global traffic for different values of MTU.

D. Energy consumption estimation as a function of MTU

After being able to obtain the global traffic for a given MTU, the energy consumption will be estimated now as a function of MTU based on: the generated traffic (part C), the results for the transmitted packets (part B.2), and the energy consumed (part B.3). This estimation can be calculated by using Eq. (6).

Table 6 and Fig. 10 show the energy consumption estimation for different values of MTU after applying Eq. (6) to the results obtained. From Fig. 10, a linear equation can be obtained for the consumed energy in the function of MTU: (8)${\displaystyle E\left(m\right)=-1.852\ast 10^{-5}\ast m+0.3195.}$ Finally, with the help of Eq. (8), the energy consumed in transmission and reception by the machine can be estimated, whatever the MTU size was in the network.