A B S T R A C T I. INTRODUCTION

Transcription

1 Building Supercomputer With Raspberry Pi Akshay Shipurkar Dilkap Research Institute Of Engineering & Management Studies,Neral University Of Mumbai, A B S T R A C T The Raspberry Pi is a credit-card sized computer that plugs into your TV and a keyboard. It has little computational power compared to today s modern day computers. It runs a striped down version of Linux.We are using it to build a supercomputer. Although it may not be as fast todays Supercomputers, But it will help to demonstrate the working of supercomputers and the concept of distributed and parallel Computing to the students and anyone with interest to build a small working Supercomputer. Although some colleges and universities have access to parallel computing hardware, none that we are aware of can provide dedicated parallel computing hardware to each student. Institutes often provide shared parallel computing equipment for the students, if they can afford to. (Abstract) Index Terms: Introduction; Raspberry pi; Raspberry pi Clusters; Related Work; Creating The Raspberry Pi Supercomputer; Raspberry Pi Setup; Building MPI To run code on multiple nodes; Cluster Results; Future Work; I. INTRODUCTION Raspberry Pi computers that cost $35 each to create a low cost supercomputer capable of distributing the processing and storage across the affordable small devices. Learning parallel computing can be challenging for students for a number of reasons. While many universities and a number of colleges teach courses on parallel computing, we are not aware of any of these institutions that have dedicated compute clusters for each student. One of the challenges students encounter when learning how to do parallel programming is getting a good understanding of how well their parallel algorithms to solve problems work. This problem is aggravated, as many students try to do their homework at the last minute, and thus, many try to use the shared cluster that their institution gives them access to at once. This can lead to small clusters operating slowly or even crashing, thus making performance testing of programs inaccurate at best, or impossible in the worst case. Although institutions with sufficient IT support likely configure their cluster to run a batch system to ensure the performance tests are accurate and can ensure that the clusters stay up constantly, this is not always feasible at smaller institutions.[1] Raspberry Pi The Raspberry Pi is a credit-card sized computer that plugs into your TV and a keyboard. It is a capable little computer which can be used in electronics projects, and for many of the things that your desktop PC does, like spreadsheets, word-processing and games. It also plays high-definition video. We want to see it being used by kids all over the world to learn programming [6] , IJAFRSE and VIVRUTI 2015 All Rights Reserved

2 . Figure 1. Raspberry Pi Model B Graphical illustration with components and specifications [7] Raspberry Pi Clusters Various groups have built Raspberry Pi clusters, we focus here on ones that were reported with HPL as a benchmark, or else have large numbers of nodes. None of them are instrumented for per-node power measurements like ours is. Pfalzgraf and Driscoll [28] create a a 25-node Raspberry Pi cluster, but do not provide power or FLOPS results. Kiepert [20] builds a 32-node Raspberry Pi cluster. He includes total power usage of the cluster, but does not include floating point performance results. Tso et al. [35] build a 56-node Raspberry Pi cloud cluster. Cox et al. [8] construct a 64-node Raspberry Pi cluster. The obtain a peak performance of 1.14 GFLOPS, which is much less than we find with 32 nodes on our cluster. Abrahamsson et al. [2] built a 300-node Raspberry Pi cluster. II. RELATED WORK Figure 2. Raspberry Pi Cluster Network Architecture [8] This project is inspired by several other projects, where the computing clusters were designed for student work especially at institutes with little money to spend on hardware. Microwulf was a Beowulf cluster built by Joel Adams and Timothy Brom that cost $2470 in 2006 [2]. Microwulf had eight CPU , IJAFRSE and VIVRUTI 2015 All Rights Reserved

3 cores and 8 GB of RAM and by 2007, its components cost only $1263 [2]. However, the Microwulf was 11"x12"x17", which is roughly the size of a desktop computer, making it more portable than other systems, but still not ideal to move around on a daily basis. The LittleFe cluster is bigger than the Microwulf, weighing in at just under 50 pounds including its traveling case [3]. The LittleFe consists of 12 CPU cores and 12 GB of RAM split between 6 nodes and includes GPGPU capabilities on each node [4]. The LittleFe costs less than $3000 and is even bigger than the Microwulf, making it more challenging to transport back and forth on a daily basis [4]. Both the Microwulf and the LittleFe are able to run programs with OpenMP and MPI since they are multi-node systems with multiple CPU cores on each node. In contrast, the San Diego Supercomputing Center's Raspberry Pi cluster, Meteor, can only run MPI programs, since the Raspberry Pi computers have only single core CPUs [5]. However, Meteor is composed of 16 Raspberry Pi computers and is more lightweight and portable than the Microwulf and LittleFe, as well as being significantly cheaper. III. CREATING THE RASPBERRY PI SUPERCOMPUTER There are two approaches to apply parallel computation for creating a supercomputer. The first approach is the shared-memory approach; the second one is the message passing interface (MPI) approach. We will be using MPI approach for making the raspberry pi supercomputer. Figure 2. Cluster design [8] Figure 3. A typical Raspberry Pi Cluster [9] IV. RASPBERRY PI SETUP 1. Get raspbian.zip image originally used: wheezy-raspbian.zip [11] It is advised to check the downloads page on raspberrypi.org and use the latest version. 2. Use win32 disk imager to put image onto an SD Card (or on a Mac e.g. Disk Utility/ dd) use the Write option to put the image from the disk to your card[12] 3. Boot on Pi 4. Expand image to fill card using the option on screen on first boot. If you don t do this on first boot, then you need to use $ sudo raspi-config 5. Log in and change the password [13] $ passwd 6. Log out and check that you typed it all OK (!) , IJAFRSE and VIVRUTI 2015 All Rights Reserved

4 $ exit 7. Log back in again with your new password V. BUILDING MPI TO RUN CODE ON MULTIPLE NODES 8. Refresh list of packages in your cache $ sudo apt-get update 9. sudo apt-get upgrade 10. Get Fortran $ sudo apt-get install gfortran 11. Make a directory to put the sources in $ mkdir /home/pi/mpich2 $ cd ~/mpich2 12. Get MPI sources from Argonne. $ wget Unpack them. $ tar xfz mpich p1.tar.gz [Note: update this as the version of MPICH2 increments] 14. $ sudo mkdir /home/rpimpi/ $ sudo mkdir /home/rpimpi/mpich2-install 15. Make a build directory (so we keep the source directory clean of build things) mkdir /home/pi/mpich_build 16. Change to the BUILD directory $ cd /home/pi/mpich_build 17. Now we are going to configure the build $ sudo /home/pi/mpich2/mpich p1/configure -prefix=/home/rpimpi/mpich2-install [Note: update this as the version of MPICH2 increments] Make a cup of tea 18. Make the files $ sudo make Make another cup of tea 19. Install the files $ sudo make install 20. Add the place that you put the install to your path $ export PATH=$PATH:/home/rpimpi/mpich2-install/bin , IJAFRSE and VIVRUTI 2015 All Rights Reserved

5 Note to permanently put this on the path edit.profile $nano ~/.profile and add at the bottom these two lines: # Add MPI to path PATH="$PATH:/home/rpimpi/mpich2-install/bin" 23. Check whether things did install or not $ which mpicc $ which mpiexec 24. Change directory back to home and create somewhere to do your tests $ cd ~ $ mkdir mpi_testing $ cd mpi_testing 25. Now we can test whether MPI works for you on a single node mpiexec -f machinefile -n <number> hostname where machinefile contains a list of IP addresses (in this case just one) for the machines a) Get your IP address $ ifconfig b) Put this into a single file called machinefile 26. $ nano machinefile Add this line: [or whatever your IP address was] 27. If you use $ mpiexec -f machinefile n 1 hostname Output is: raspberrypi 28. Now to run a little C code. In the examples subdirectory of where you built MPI is the famous CPI example. You will now use MPI on your Pi to calculate Pi: $ cd /home/pi/mpi_testing $ mpiexec -f machinefile -n 2 ~/mpich_build/examples/cpi Output is Process 0 of 2 is on raspberrypi Process 1 of 2 is on raspberrypi pi is approximately , Error is Celebrate if you get this far , IJAFRSE and VIVRUTI 2015 All Rights Reserved

6 Flash me once International Journal of Advance Foundation And Research In Science & Engineering (IJAFRSE) 29. We now have a master copy of the main node of the machine with all of the installed files for MPI in a single place. We now want to clone this card. 30. Shutdown your Pi very carefully $ sudo poweroff Remove the SD Card and pop it back into your SD Card writer on your PC/ other device. Use Win32 disk imager (or on a Mac e.g. Disk Utility/ dd) to put the image FROM your SD Card back TO your PC: You will use the Read option to put the image from the disk to your card Let us call the image wheezy-raspbian_backup_mpi_master.img 31. Eject the card and put a fresh card into your PC/other device. Use win32 disk imager to put image onto an SD Card (or on a Mac e.g. Disk Utility/ dd) You will use the Write option to put the image from the disk to your card and choose the wheezy-raspbian_backup_mpi_master.img image you just created. [Note that there are probably more efficient ways of doing this in particular maybe avoid expanding the filesystem in step 4 of the first section.] 32. Put the card into your second Pi and boot this. You should now have two Raspberry Pis on. Unless otherwise stated, all the commands below are typed from the Master Pi that you built first. Using SSH instead of password login between the Pis 33. Sort out RSA to allow quick log in. This is the best thing to read: In summary (working on the MASTER Pi node) $ cd ~ $ ssh-keygen -t rsa C raspberrypi@raspberrypi This set a default location of /home/pi/.ssh/id_rsa to store the key Enter a passphrase e.g. myfirstpicluster. If you leave this blank (not such good security) then no further typing of passphrases is needed. $ cat ~/.ssh/id_rsa.pub ssh pi@ "mkdir.ssh;cat >>.ssh/authorized_keys" 34. If you now log into your other Pi and do $ ls al ~/.ssh You should see a file called authorized_keys this is your ticket to no login heaven on the nodes 35. Now let us add the new Pi to the machinefile. (Log into it and get its IP address, as above) Working on the Master Raspberry Pi (the first one you built): $ nano machinefile Make it read , IJAFRSE and VIVRUTI 2015 All Rights Reserved

7 International Journal of Advance Foundation And Research In Science & Engineering (IJAFRSE) [The two IP addresses you have for the machines are] 36. $ cd /home/pi/mpi_testing $ mpiexec -f machinefile -n 2 ~/mpich_build/examples/cpi Output is Process 0 of 2 is on raspberrypi Process 1 of 2 is on raspberrypi pi is approximately , Error is If changing the hostname on second machine (see Appendix 1 Hostname Script ) and run: $ mpiexec -f machinefile -n 2 ~/mpich_build/examples/cpi Output: Process 0 of 2 is on raspberrypi Process 1 of 2 is on iridispi002 Now you can see each process running on the separate nodes. VI. CLUSTER RESULTS The HPL benchmark on our Raspberry Pi cluster while measuring the power consumption in order to calculate FLOPS per Watt metrics Peak FLOPS results The peak performance of our Pi cluster running with 32 boards (16 Model B, 16 Model B+). We find a peak of 4.37 GFLOPS while using 93.0W for a MFLOPS/W rating of If overclocked the cluster to run at 1GHz[10] Figure 4. MFLOPS with number of nodes for Raspberry Pi cluster, no overclocking [10] , IJAFRSE and VIVRUTI 2015 All Rights Reserved

8 Figure 5. Average power with number of nodes for Raspberry Pi cluster, no overclocking. The dropoff with 32 nodes is because the additional 16 nodes use Model B+ (instead of Model B) which use less power [10]. VII. FUTURE WORK Expand the size. We have parts to expand to 48 nodes and can easily expand to 64 and beyond. Enable hardware performance counter support. Two of the authors have gotten Raspberry Pi performance counter support merged into upstream Linux, but this has not made it to distributions such as Raspbian yet. Having access to the counters in conjunction with the energy measurements will enable more thorough performance studies. Harness the GPUs. Table 2 shows the GPU capabilities available on the various boards. The Raspberry Pi has a potential 24 GFLOPS available perf node which is over an order of magnitude more than found on the CPU. Grasso et al. [15] use OpenCL on a Cortex A15 board with a Mali GPU and find they can get 8.7 times better performance than the CPU with 1/3 the energy. If similar work could be done to obtain GPGPU support on the Raspberry Pi our cluster could obtain a huge performance boost. VIII. REFERENCES [1] Toth, David. "A Portable Cluster for Each Student." Parallel & Distributed Processing Symposium Workshops (IPDPSW), 2014 IEEE International. IEEE, [2] J. Adams and T. Brom, "Microwulf: a beowulf cluster for every desk," Proc. 39th SIGCSE technical symposium on Computer science education (SIGCSE '08), Mar. 2008, pp , doi: / [3] C. Peck, "LittleFe - parallel and distributed computing education on the move," Journal of Computing Sciences in Colleges, vol. 26, Oct. 2010, pp [4] Welcome to LittleFe.net LittleFe , IJAFRSE and VIVRUTI 2015 All Rights Reserved

9 [5] SDSC Uses Meteor Raspberry Pi Cluster to Teach Parallel Computing l_computing. [6] Available at: [7] Available at: [8] Franks, Shaun, and Johnathan Yerby. "CREATING A LOW-COST SUPERCOMPUTER WITH RASPBERRY PI." [9] Available at: [10] Cloutier, Michael F., Chad Paradis, and Vincent M. Weaver. "Design and Analysis of a 32-bit Embedded High-Performance Cluster Optimized for Energy and Performance Extended Edition." [11] Available at: [12] Available at: Imager.shtml [13] Available at: AUTHORS PROFILE Mr.Akshay Shipurkar He is currently persuing his Bachlors Degree in Computer Engineering from Dilkap Research Institute Of Engineering & Management Studeis, Neral (University Of Mumbai) and is currently in third year. His intreasts include Super computing, Computer networks, Cryptography. He is currently intrim IEEE Student Chair of IEEE DRIEMS and likes to participate in technical compitions , IJAFRSE and VIVRUTI 2015 All Rights Reserved