Trends for adapting High performance computing

With HPC, the primary goal is to crunch numbers, not to sort data. It demands specialized program optimizations to get the most from a system in terms of input/output, computation, and data movement. And the machines all have to trust each other because they’re shipping information back and forth.

HPC systems are based on parallel computing, which uses concurrency to decrease program runtimes. If the program lends itself to parallelization, the more nodes you can throw at it, the better. It’s like painting a fence; if you can get some friends to paint with you, the job gets done faster. But some applications must follow a serial process, like building a house. You may have an application that can only run on one CPU, with no shortcutting, no way to split it up. In that case, you could just get a bigger, faster box.

In the aerospace industry you’ve got a lot of calculations to do before you can call an airplane design good and ship it. Or if you’re in the oil industry, there may be a patch of ground somewhere that looks promising, but the geological information requires more processing. So you take the information from the field back and throw it at the cluster computer, and run it through some powerful software to figure out if the land has oil. First one to figure if it does, gets the bid in and wins. The other guy loses.

n HPC clusters, all of the machines are working on a specific numeric problem that needs a significant amount of calculation to find the answer. For example, although it’s really fast, a state DMV driver license database lookup is not HPC, it’s more of a table lookup, not a calculation. But if you’re building airplanes and calculating the airflow over a wing of a 747, the more speed the better. You have to break up the wing sections into one inch pieces, then calculate the flow over each sq. inch, etc. then repeat the process. The faster the cluster, the faster you’re going to get your answer.

There are three rules when building HPC clustering. Rule number one is, the application defines how the cluster is built: everything underneath it is dependent on that application. Just because a processor is faster doesn’t mean it’s going to be faster on that application.

Rule two is the balance between computing and communication. A cluster’s individual members compute like crazy, then at some point they communicate. Then it repeats; it’s another difference in how you design your hardware. With a fixed amount of money to work from, you have to split it between hardware for communication, and the hardware for computation or nodes, infrastructure, etc.

The third rule of clustering is “There will be an upgrade”! Clusters are not usually static, they can grow machine by machine as the computing needs of the organization change