How to Build a Supercomputer

An easy-to-follow guide on making a supercomputer

Are you in need of a machine that can deliver hundreds of trillions of floating-point calculations per second? Or are you in need of a bar story about how the supercomputer in your basement flipped a breaker? Building your own High-Performance Compute cluster, a.k.a. supercomputer, is a challenge any expert geek with a weekend of free time and some cash to burn can tackle. Technically speaking, a modern, multi-processor supercomputer is a network of computers working together in parallel to solve a problem. This article will briefly describe each step in the process, focusing on hardware and software.

Steps

Download Article

1
First determine the hardware components and resources needed. You will need one head node, at least a dozen identical compute nodes, an Ethernet switch, a power distribution unit, and a rack. Determine the electrical demand, cooling, and space required. Also decide on what IP address you want for your private networks, what to name the nodes, what software packages you want installed, and what technology you want to provide the parallel computing capabilities (more on this later).
- Though the hardware is expensive, all software listed in this how-to is free, and most are open source.
- If you would like to see how fast your supercomputer would theoretically be, use this tool: http://hpl-calculator.sourceforge.net/
2
Build the compute nodes. You will need to assemble the compute nodes or acquire pre-build servers.
- Choose a computer server chassis that maximizes space, cooling, and energy efficiency.
- Or you can utilize a dozen or so used, outdated servers - whose whole will outweigh the sum of their parts yet save you a sizable lump of cash. All processors, network adapters, and motherboards should be identical for the whole system to play together nicely. Of course, don't forget about RAM and storage for each node and at least one optical drive for the head node.
Advertisement
Start from the bottom, so the rack isn't top heavy. You will need a friend to help you with this - the dense servers can be very heavy and guiding them into the rails that hold them into the rack is difficult.
Take this moment to configure the switch: allow for jumbo frame sizes of 9000 bytes, set the IP address to the static address you decided on in step 1, and turn off unnecessary routing protocols such as SMTP Snooping.
Depending on how much current your nodes may need at maximum load, you may need 220 volts for high performance computing.
6
With everything installed, you can begin the configuration process. Linux is the de facto OS for HPC clusters–not only is it the ideal environment for scientific computing, but it doesn't cost a thing to install it on hundreds or even thousands of nodes. Imagine how much it would cost to install Windows on all those nodes!
- Begin with installing the latest version of the motherboard BIOS and firmware, which should be the same on all nodes.
- Install your preferred linux distro on each node, with a graphical UI for the head node. Popular choices include CentOS, OpenSuse, Scientific Linux, RedHat, and SLES.
- This author highly recommends using the Rocks Cluster Distribution. In addition to installing all the tools necessary for a compute cluster to function, Rocks uses a great method for 'distributing' many instances of itself to the nodes very quickly using PXE boot and the Red Hat 'Kick Start' procedure.
7
Install the message-passing interface, resource-manager, and other necessary libraries. If you didn't install Rocks in the previous step, you will have to manually setup the necessary software to enable the parallel computing mechanisms.
- First you will need a portable bash management system, such as the Torque Resource Manager, which allows you to break-up and distribute tasks to multiple machines.
- Pair Torque with the Maui Cluster Scheduler to complete the setup.
- Next you will need to install the message passing interface, necessary for the individual processes on the separate compute nodes to share the same data. OpenMP is a no-brainer.
- Don't forget the multi-threading math libraries and compilers to build your parallel computing programs. Did I mention that you should just install Rocks?
8
Network the compute nodes together. The head node sends the compute tasks to the compute nodes, which in turn must send the result back, as well as sending messages to each other. The faster the better.
- Use a private ethernet network to connect all the nodes in the cluster.
- The head node can also act as a NFS, PXE, DHCP, TFTP, and NTP server over the Ethernet network.
- You must separate this network from public networks, which ensures that broadcast packets don't interfere with other networks in your LAN.
9
Test the cluster. The last thing you may want to do before releasing all this compute power to your users is test it's performance. The HPL (High Performance Lynpack) benchmark is a popular choice for measuring the computational speed of the cluster. You will need to compile it from source with all possible optimizations your compiler offers for the architecture you chose.
- You must, of course, compile from source with all possible optimization options for your platform. For example, if using AMD CPUs, compile with Open64 with -0fast optimization level.
- Compare your results on TOP500.org to compare your cluster to the fastest 500 supercomputers in the world!

Community Q&A

Search

Add New Question

Question
Can I link inexpensive, low-power Windows 10 desktops together to make a "mini" supercomputer?
Community Answer
It is certainly possible to link a number of ordinary PCs together to make a small cluster. However, I do not think that you would be able to do this easily using the existing Windows operating systems as I don't think anyone has really pursued this (probably due to reasons of cost and proprietary issues). You could, however, easily use the same PCs, without affecting the existing O/S either by installing Linux as a "dual boot" system or installing it on multiple USB memory sticks and booting the PCs from them. Linux is really the only choice for doing this as it is the system on which most work has been done in this field, with some versions being specifically designed for clusters.
Thanks! We're glad this was helpful.
Thank you for your feedback.
If How.com.vn has helped you, please consider a small contribution to support us in helping more readers like you. We’re committed to providing the world with free how-to resources, and even $1 helps us in our mission.Support How.com.vn
Yes No
Not Helpful 7 Helpful 33
Question
Can I use Raspberry Pi as compute nodes when building a supercomputer?
Community Answer
Yes, but don't expect very good performance. Raspberry Pi's aren't very powerful, but they are silent and output almost no heat, which is ideal.
Thanks! We're glad this was helpful.
Thank you for your feedback.
If How.com.vn has helped you, please consider a small contribution to support us in helping more readers like you. We’re committed to providing the world with free how-to resources, and even $1 helps us in our mission.Support How.com.vn
Yes No
Not Helpful 4 Helpful 18
Question
How hot can a supercomputer get?
Community Answer
It really depends on how the power distribution is set up. Too much power can heat it up much more. Not having a good cooling method can also cause heat. Make sure you have a good power distribution method, even if you have to pay more -- it is better than short circuiting your nodes.
Thanks! We're glad this was helpful.
Thank you for your feedback.
If How.com.vn has helped you, please consider a small contribution to support us in helping more readers like you. We’re committed to providing the world with free how-to resources, and even $1 helps us in our mission.Support How.com.vn
Yes No
Not Helpful 7 Helpful 22

See more answers

Ask a Question

200 characters left

Include your email address to get a message when this question is answered.

Submit

Video

Tips

For really high network speeds, look into the InfiniBand network interfaces. Be prepared to pay premium prices though.
Thanks
Helpful17 Not Helpful12
IPMI can make administration of a large cluster a breeze by providing KVM-over-IP, remote power cycling, and more.
Thanks
Helpful14 Not Helpful14
Use Ganglia to monitor the computational loads on the nodes.
Thanks
Helpful9 Not Helpful17