First, determine the required hardware components and resources. You will need a head node with at least a dozen identical compute nodes, Ethernet switches, power distribution equipment and racks. Clear power demand, cooling and space requirements. Also decide whether you want to name your private IP address as a node, the software package you want to install, and what kind of technical parallel computing capability you want to provide (more about this version). Although the hardware can't be downloaded at will, the listed software is free and most of them are open source. (2)/ Establish a computing node. You will need to assemble compute nodes or get pre-built servers. More computing nodes and more powerful supercomputers! Choose a server chassis to maximize space, cooling and energy efficiency. Or, you can use more than a dozen such outdated servers-the whole will exceed the sum of its parts and save you considerable one-time cash. Play the whole system well, all processors, network cards and motherboards should be the same. Of course, don't forget the memory and the head node that stores each node and at least one optical drive.

3. The server is installed on the rack. Starting from the bottom, the shelf is not the heaviest. You need a friend to help you-high-density servers can be very heavy and it is difficult to guide them to rack rails.

4. Ethernet switch installed in the above server chassis. Use this moment to configure the switch: allow a huge frame size of 9000 bytes, set the static address of the IP address you decided in step 1, and turn off unnecessary routing protocols, such as SMTP interception.

5. PDU (power distribution device) has been installed. You may need to calculate high performance according to how much current your node may need under the maximum load of 220 volts.

Once everything is installed, you can start the configuration process. Linux is the de facto operating system. HPC cluster is not only an ideal environment for scientific computing, but also can be installed on hundreds or even thousands of nodes without any cost. Imagine the cost of installing Windows on all these nodes. First, install the latest version of the motherboard BIOS and firmware, which should be the same on all nodes. Your preferred Linux distribution is installed on each node, which is the head node of the graphical user interface. Popular choices include CentOS, OpenSUSE, Red Hat Linux and SLES. In this paper, the author strongly recommends the use of rock cluster distribution. In addition to installing all necessary tools and calculating cluster functions, the nodes of the rock itself are very fast in many cases, so it is a good method to start "program" scheduling by using PXE boot and Red Hat's feet.

7. Install the messaging interface, resource manager and other necessary libraries. If you didn't install rock in the previous step, you will have to manually set up the necessary software to make the parallel computing mechanism. First of all, you need a portable bash management system, such as Torque resource manager, which allows you to decompose and distribute tasks to multiple machines. The installation of torque Maui cluster scheduling has been completed. Next, you need to install the messaging interface, and all processes on the independent computing node must * * * enjoy the same data. OpenMP is a brainless person. Don't forget the multithreaded math library and compiler to build parallel computing programs. Did I mention that you should only install rocks?

8. Network computing nodes. The computing node of the head node must send the returned results and messages to the other party to send the computing task. The sooner the better. Connect the wires! Connect all nodes in the cluster using private Ethernet. The head node can also act as NFS, PXE, DHCP, TFTP and NTP servers through Ethernet. From the public * * * network, in order to ensure that broadcast packets will not interfere with other networks, you must connect to this network in your local area network.

9. Test the cluster. Before you do something in the past, you may want to release the computing power of all users to test its performance. The computing speed of HPL (High Performance Lynpack) benchmark cluster is a popular choice. You need to compile from the source code, and all possible optimization compilers provide the architecture of your choice. Of course, you must use all possible optimization options to compile from the source code of your platform. For example, if you use AMD's CPU, compile with 0fast optimization level of Open64. The world's fastest 500 supercomputer clusters compare TOP500, and compare the results!

point out

Use Ganglia to monitor the computing load on the node. IPMI can provide KVM-over-IP, remote power cycle and a larger cluster management. For very high network speed, interface to InfiniBand network. Ready, but at a higher price.