Why supercomputing?

If supercomputing is that complex, why would one want to use it?

When a PC or server fails...

Processing of large datasets may require more storage than a workstation or a departmental server can deliver, or it may require more bandwidth to memory or disk than a workstation or server can deliver, or simply more processing power than a workstation can deliver. In many cases, a supercomputer may help in those cases. However, it may require different or reworked software.

Large simulations, e.g., partial differential equations in various fields of physics, may require far more processing power than a workstation can deliver, and often result in large datasets (which then brings us back to the previous point). Supercomputers are often extremely suited for those very large simulations, but again it may require different software or a significant reworking of the software.

But there is also an easier case: applications such as parameter analysis or Monte Carlo sampling. In these cases we may want to run thousands of related simulations with different input. A workstation may have more than enough processing power to process a single or a few samples, but to work on thousands of those we need a lot more processing power if we want to finish the work in a reasonable amount of time. Here also supercomputer can come to the rescue.

Supercomputing jobs

From the above discussion we already get the feeling that there are two big reasons to use a supercomputer.

1. We may want to use a supercomputer simply because we want to improve the turnaround time for a large computation, or because the computation is not even feasible on a smaller computer because of the required compute time and memory capacity. This is also called capability computing. In capability computing one typically thinks in terms of hours per job.

2. We may also want to use a supercomputer to improve throughput if we have to run a lot of smaller jobs. This is called capacity computing. In capacity computing one typically thinks in terms of jobs per hour.

Furthermore, we can distinguish simulation and data processing as two big domains where supercomputing could be used. Some will add AI as a third pillar, but AI is typically used for data processing and has requirements similar to the other data processing jobs we will discuss. And sometimes the finished AI model is then used as part of the model in a simulation.

Supercomputers are really build for capability computing jobs. They can also accommodate many (but often not all types of) capacity computing jobs, and often capacity computing jobs could be run equally well or better on just a network of servers, e.g., a cloud infrastructure, at possibly a much lower cost (should the user be charged a realistic amount for the compute resources consumed).

There are many examples of capability computing in simulation. Computational fluid dynamics of turbulent and even many laminar flows requires a huge amount of compute capacity. The demand becomes only higher when considering fluid-structure interactions causing, e.g., vibration of the structure. Another example is virtual crash tests during the design of cars. That a car manufacturer can produce so much more different models than 30 years ago is partly the result from the fact that developing a different body for a car is now much cheaper than it was 30 years ago despite the more stringent safety rules, as far less prototypes need to be build before having a successful crash tests thanks to the virtual crash tests. Weather and climate modeling are also examples of capability computing, as is the simulation of large complex molecules and their chemical reactions.

But there are also examples of capacity computing in simulation. E.g., when doing a parameter study that requires simulating a system for multiple sets of parameters. In drug design one often wants to test a range of molecules for certain properties. The risk analysis computations done by banks are also an example of capacity computing. Sometimes one really has a combination of capacity computing and capability computing, e.g., when doing a parameter analysis for a more complex system.

Data processing can also lead to capability computing. One example is the visualisation of very large data sets or simulation results, were a visualisation workstation may not be enough. Training AI models is also data-driven and often very data-intensive. Another example are the electronic stamps offered by the US postal services, certainly in the initial years. Buyers of electronic stamps would get codes that they can each use once, but of course US postal needs to check not only if a code is valid but also if it has not been used yet. And a clever abuser may share codes with someone at the opposite site of the country. Yet US postal has to be able to detect abuse in real time to not slow down the sorting machines. For this they used a computer which around 2010 (when the author of these notes learned about this application) could truly be considered a supercomputer.

Examples of capacity computing for data processing are many of the data mining applications that often consist of many rather independent tasks. The processing of the CERN Large Hadron Collider data is also mostly capacity computing (and one that doesn't always work well on large supercomputers). Another example is a researcher from a Flemish university who used a VSC supercomputer to pre-process a text corpus consisting of newspaper articles of many years. All texts had to be grammatically analysed. After the preprocessing, a regular server was enough for the research.