Academia.eduAcademia.edu

Figure 11 – uploaded by Dmytro Zubov

See full PDFdownloadDownload figure
Fig. 13. Building the installation ISO file.

Figure 13 Building the installation ISO file.

Related Figures (16)

Fig. 1. An example of the output of two Python programs that calculate prime numbers.
Fig. 1. An example of the output of two Python programs that calculate prime numbers.
Fig. 3. MPICH heterogeneous cluster: Raspberry Pi rack mount and router D-Link DIR-300/ with a built-in 4-port switch. SUA Ve Be eS Ee The Raspberry Pi rack mount and router D-Link DIR-300A with a built-in 4-port switch are shown in Fig. 3. The naster, second, and third nodes had IP addresses of 192.168.0.92, 192.168.0.41, and 192.168.0.13, respectively. Boards are connected only by Ethernet cables via the router D-Link DIR-300A and the power supply, because they are remotely controlled by the SSH protocol. In addition, a virtual network computing (VNC) server was installed on the master node to control the Raspberry Pi 4B through RealVNC remote access software [22]. The cluster had a two-dimensional mesh topol- ogy [23]. The execution times for one (data series 1, performance weights (1, 0, 0), one Rasp- berry Pi 4B board), two (data series 2, performance weights (1, 1, 0), two Raspberry Pi 4B boards), and three (data series 3, performance weights (1, 1, 1), two 4B and one 3B Raspberry Pi boards) cluster nodes are presented in Fig. 4 for different values of the variable end_number, i.e., the maximum integer number employed in the current it- eration. Here, the execution time for a specific point is the average of three appropriate values. A maximum of two MPICH commands mpiexec [4] and an additional Python
Fig. 3. MPICH heterogeneous cluster: Raspberry Pi rack mount and router D-Link DIR-300/ with a built-in 4-port switch. SUA Ve Be eS Ee The Raspberry Pi rack mount and router D-Link DIR-300A with a built-in 4-port switch are shown in Fig. 3. The naster, second, and third nodes had IP addresses of 192.168.0.92, 192.168.0.41, and 192.168.0.13, respectively. Boards are connected only by Ethernet cables via the router D-Link DIR-300A and the power supply, because they are remotely controlled by the SSH protocol. In addition, a virtual network computing (VNC) server was installed on the master node to control the Raspberry Pi 4B through RealVNC remote access software [22]. The cluster had a two-dimensional mesh topol- ogy [23]. The execution times for one (data series 1, performance weights (1, 0, 0), one Rasp- berry Pi 4B board), two (data series 2, performance weights (1, 1, 0), two Raspberry Pi 4B boards), and three (data series 3, performance weights (1, 1, 1), two 4B and one 3B Raspberry Pi boards) cluster nodes are presented in Fig. 4 for different values of the variable end_number, i.e., the maximum integer number employed in the current it- eration. Here, the execution time for a specific point is the average of three appropriate values. A maximum of two MPICH commands mpiexec [4] and an additional Python
Fig. 4. Execution times for one (data series 1), two (data series 2), and three (data series 3) cluster nodes for different values of the maximum integer number: Basic implementation of the prime numbers finding algorithm [4]. program were run in parallel in different terminals on Raspberry Pi 4B board one MPICH Python program was run on a Raspberry Pi 3B board to avoid overloading and overheating the CPU and RAM. Analysis of the data series presented in Fig. 4 shows that the basic implementation of the prime numbers finding algorit lowest performance with three nodes. This result does not contrad s, and only hm has the ict Amdahl's law [23] because the third node with the Raspberry Pi 3B board is significantly slower than the first and second nodes. The fluctuations in the data series can be linear CPU speed-temperature dependencies of Raspberry Pi explained boards [24 by the non- ], different Ethernet connectivity (maximum throughput 300 Mbps for 3B and | Gbps for 4B), and unstable and low-speed network bandwidth, which lead to substantial performance deg- radation in the grid environment [25, 26].
Fig. 4. Execution times for one (data series 1), two (data series 2), and three (data series 3) cluster nodes for different values of the maximum integer number: Basic implementation of the prime numbers finding algorithm [4]. program were run in parallel in different terminals on Raspberry Pi 4B board one MPICH Python program was run on a Raspberry Pi 3B board to avoid overloading and overheating the CPU and RAM. Analysis of the data series presented in Fig. 4 shows that the basic implementation of the prime numbers finding algorit lowest performance with three nodes. This result does not contrad s, and only hm has the ict Amdahl's law [23] because the third node with the Raspberry Pi 3B board is significantly slower than the first and second nodes. The fluctuations in the data series can be linear CPU speed-temperature dependencies of Raspberry Pi explained boards [24 by the non- ], different Ethernet connectivity (maximum throughput 300 Mbps for 3B and | Gbps for 4B), and unstable and low-speed network bandwidth, which lead to substantial performance deg- radation in the grid environment [25, 26].
Fig. 5. Raspberry Pi command top: An example of the processes currently running on the cluster nodes with the equally distributed workload. generates fuzzy singleton values [19]. Many other defuzzification methods have been proposed, but none of them generates an output for the common sense of knowledge modeling [19]. The defuzzification process does not have a uniform framework [19]. In addition, the computational results of defuzzification are often in conflict [13-18].
Fig. 5. Raspberry Pi command top: An example of the processes currently running on the cluster nodes with the equally distributed workload. generates fuzzy singleton values [19]. Many other defuzzification methods have been proposed, but none of them generates an output for the common sense of knowledge modeling [19]. The defuzzification process does not have a uniform framework [19]. In addition, the computational results of defuzzification are often in conflict [13-18].
rable 1. Linguistic and numeric variables zj and weights ai for Raspberry Pi MPIC heterogeneous cluster soft-/hardware Table 2. Assessment of performance weights for Raspberry Pi MPICH cluster nodes
rable 1. Linguistic and numeric variables zj and weights ai for Raspberry Pi MPIC heterogeneous cluster soft-/hardware Table 2. Assessment of performance weights for Raspberry Pi MPICH cluster nodes
Fig. 6. Pseudocode of the prime numbers finding algorithm with weight-based load balancing in Raspberry Pi MPICH heterogeneous cluster.
Fig. 6. Pseudocode of the prime numbers finding algorithm with weight-based load balancing in Raspberry Pi MPICH heterogeneous cluster.
Fig. 7. Weight-based load balancing implementation of the prime numbers finding algorithm: Execution times for different values of the maximum integer number and performance weights (4, 4, 1), , 5, 1), and (6, 6, 1). The developed Python program was executed on the Raspberry Pi with the aforementioned three nodes. The execution times for perfo MPICH cluster rmance weights (4, 4, 1), (5, 5, 1), and (6, 6, 1) are presented in Fig. 7 for different values of the variable end number: dash, grey, and black lines, respectively. Here, the ex ecution time for a specific point is the average of three appropriate values too. In addition, a maximum of two MPICH commands mpiexec and an additional Python program were run in parallel in different terminals on Raspberry Pi 4B boards, and only one program was run on a Raspberry Pi 3B board to avoid overloading and MPICH Python overheating the CPU and RAM. Analysis of the data series presented in Fig. 7 shows that the mean signed deviation for data series with performance weights (5, 5, 1)-(6, 6, 1) is -3.01 s and (5,5, 1)-(4, 4, 1) is -155.74 s, i.e., data series with recommend weights (5, 5, 1) has less average execution time across the entire set of In addition, a three-tuple (5, 5, 1) is preferable compared with (6, 6, ed performance all observations. ) because it has less workload for the first and the second nodes along with a greater workload for the third node, which corresponds to Table 2. This result can be explained timeout of the third node with the low-performance Raspberry Pi 3B by the minimum board. Fluctua- tions in the data series appear because of the unstable network bandwidth and non- linear CPU speed-temperature dependencies of the Raspberry Pi board s too.
Fig. 7. Weight-based load balancing implementation of the prime numbers finding algorithm: Execution times for different values of the maximum integer number and performance weights (4, 4, 1), , 5, 1), and (6, 6, 1). The developed Python program was executed on the Raspberry Pi with the aforementioned three nodes. The execution times for perfo MPICH cluster rmance weights (4, 4, 1), (5, 5, 1), and (6, 6, 1) are presented in Fig. 7 for different values of the variable end number: dash, grey, and black lines, respectively. Here, the ex ecution time for a specific point is the average of three appropriate values too. In addition, a maximum of two MPICH commands mpiexec and an additional Python program were run in parallel in different terminals on Raspberry Pi 4B boards, and only one program was run on a Raspberry Pi 3B board to avoid overloading and MPICH Python overheating the CPU and RAM. Analysis of the data series presented in Fig. 7 shows that the mean signed deviation for data series with performance weights (5, 5, 1)-(6, 6, 1) is -3.01 s and (5,5, 1)-(4, 4, 1) is -155.74 s, i.e., data series with recommend weights (5, 5, 1) has less average execution time across the entire set of In addition, a three-tuple (5, 5, 1) is preferable compared with (6, 6, ed performance all observations. ) because it has less workload for the first and the second nodes along with a greater workload for the third node, which corresponds to Table 2. This result can be explained timeout of the third node with the low-performance Raspberry Pi 3B by the minimum board. Fluctua- tions in the data series appear because of the unstable network bandwidth and non- linear CPU speed-temperature dependencies of the Raspberry Pi board s too.
Fig. 8. Raspberry Pi command top: An example of the processes currently running on the cluster nodes with weight-based load balancing.
Fig. 8. Raspberry Pi command top: An example of the processes currently running on the cluster nodes with weight-based load balancing.
Fig. 9. Hardware complex of MPICH cluster on hypervisor VMware ESXi ARM Fling: Rasp- berry Pi rack mount and router TP-Link TL-WR940N. la The hardware complex consists of a Raspberry Pi rack mount with two Raspberry Pi 4B 8 GB boards and a router TP-Link TL-WR940N with four Ethernet ports (see Fig. 9). Raspberry Pi 64-bit OSs were executed using two microSD cards of 8 GB each. The hypervisor VMware ESXi ARM Fling with build number 17230755 [20] was in- stalled on a USB memory stick of 128 GB. A microSD card of 128 MB only starts the unified extensible firmware interface (UEFI) boot, and then the control is conveyed to the hypervisor VMware ESXi ARM Fling on the USB memory stick. In virtual ma- chines, the guest OS is an Ubuntu Server for ARM 20.04.2 LTS. The router operates in range extender mode. In the Raspberry Pi MPICH cluster, the master and second nodes had IP addresses 192.168.0.101 and 192.168.0.102, respectively. Boards are connected only by Ethernet cables via the router TP-Link TL-WR940N and the power supply, because the MPICH cluster is remotely controlled by SSH protocol.
Fig. 9. Hardware complex of MPICH cluster on hypervisor VMware ESXi ARM Fling: Rasp- berry Pi rack mount and router TP-Link TL-WR940N. la The hardware complex consists of a Raspberry Pi rack mount with two Raspberry Pi 4B 8 GB boards and a router TP-Link TL-WR940N with four Ethernet ports (see Fig. 9). Raspberry Pi 64-bit OSs were executed using two microSD cards of 8 GB each. The hypervisor VMware ESXi ARM Fling with build number 17230755 [20] was in- stalled on a USB memory stick of 128 GB. A microSD card of 128 MB only starts the unified extensible firmware interface (UEFI) boot, and then the control is conveyed to the hypervisor VMware ESXi ARM Fling on the USB memory stick. In virtual ma- chines, the guest OS is an Ubuntu Server for ARM 20.04.2 LTS. The router operates in range extender mode. In the Raspberry Pi MPICH cluster, the master and second nodes had IP addresses 192.168.0.101 and 192.168.0.102, respectively. Boards are connected only by Ethernet cables via the router TP-Link TL-WR940N and the power supply, because the MPICH cluster is remotely controlled by SSH protocol.
Fig. 10. Execution times for Ubuntu Server for ARM 20.04.2 LTS on VMware ESXi ARM Fling (data series 1) and Raspberry Pi 64-bit OS (data series 2) for different values of the maximum integer number: Prime numbers finding algorithm [4] in the MPICH cluster. The execution times for Ubuntu Server for ARM 20.04.2 LTS on VMware ESXi ARM Fling (data series 1) and Raspberry Pi 64-bit OS (data series 2) are presented in Fig. 10 for different values of the maximum integer number. Here, the execution time for a specific point is the average of three appropriate values. Fluctuations in data series can be explained by non-linear CPU speed-temperature dependencies of Raspberry Pi boards and unstable and low-speed network bandwidth, which lead to substantial per- formance degradation in the grid environment [25, 26].
Fig. 10. Execution times for Ubuntu Server for ARM 20.04.2 LTS on VMware ESXi ARM Fling (data series 1) and Raspberry Pi 64-bit OS (data series 2) for different values of the maximum integer number: Prime numbers finding algorithm [4] in the MPICH cluster. The execution times for Ubuntu Server for ARM 20.04.2 LTS on VMware ESXi ARM Fling (data series 1) and Raspberry Pi 64-bit OS (data series 2) are presented in Fig. 10 for different values of the maximum integer number. Here, the execution time for a specific point is the average of three appropriate values. Fluctuations in data series can be explained by non-linear CPU speed-temperature dependencies of Raspberry Pi boards and unstable and low-speed network bandwidth, which lead to substantial per- formance degradation in the grid environment [25, 26].
Fig. 11. Website https://aupdump.net with an installation image of Windows 11 ARM OS.
Fig. 11. Website https://aupdump.net with an installation image of Windows 11 ARM OS.
Fig. 12. Creating the download package with a batch file uup_download_windows.cmd.
Fig. 12. Creating the download package with a batch file uup_download_windows.cmd.
Fig. 14. Flashing the ISO file onto a microSD card using the Windows on Raspberry Imager software. the laptop Lenovo G510 with Intel Core 17-4700MQ CPU (eight logical processors — two per one physical core) and peak energy consumption of 65 W (the wattage of AC adapter is considered here), one iteration takes approximately an hour. The following ratios represent the approximate power consumed by the devices to calculate one iter- ation on one core/logical processor:
Fig. 14. Flashing the ISO file onto a microSD card using the Windows on Raspberry Imager software. the laptop Lenovo G510 with Intel Core 17-4700MQ CPU (eight logical processors — two per one physical core) and peak energy consumption of 65 W (the wattage of AC adapter is considered here), one iteration takes approximately an hour. The following ratios represent the approximate power consumed by the devices to calculate one iter- ation on one core/logical processor:
Fig. 15. Windows 11 ARM OS on Raspberry Pi 4B 8 GB: Desktop user interface with three running 32-bit Windows applications.
Fig. 15. Windows 11 ARM OS on Raspberry Pi 4B 8 GB: Desktop user interface with three running 32-bit Windows applications.
Fig. 16. Windows 11 ARM OS on Raspberry Pi 4B 8 GB: An example of the application proces- sor affinity.
Fig. 16. Windows 11 ARM OS on Raspberry Pi 4B 8 GB: An example of the application proces- sor affinity.
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved