Amdahl's law needs to be taken seriously

A computer researcher is close to rewriting Amdahl’s law, which is a theory that focuses on performance relative to parallel and serial computing.

In a presentation entitled “Breaking the Law” at the International Supercomputing Conference this week in Leipzig, Germany, Thomas Lippert will reveal how he thinks the law can be adabted.

For those who came in late, Amdahl’s law was worked out in 1967 by noted Big Blue boffin Gene Amdahl who just happened to have the same name as the law. It is supposed to explain the limitations of parallel computing based on certain models.

The theory states if you chuck enough hardware in a parallel set up you can solve a problem but you still have some in serial, which is a limiting factor in speeding up tasks.

The maths assume there is a limit to parallel speed-up, assuming some things are constant, such as the problem size and the nature of the processors doing the computation.

Amdahl’s law was re-evaluated by John Gustafson who came up with Gustafson’s law which said that if problem size is not constant and parallel computer speed can scale up accordingly.

Lippert was looking at experiments done as part of the DEEP Project, which investigates highly parallel computing models that help speed up supercomputers.

That project involves building high-performance systems called JUROPA (Jülich Research on Petaflop Architectures) and was set up by Bull, Partec and Intel.

“This machine is ideal for highly complex problems that exhibit a lower concurrency, in general. Most codes live somewhere in between. I want to find out, if we can bring the concepts together. The different architectures can assign the … different code parts according to the concurrency”, Lippert said.

Writing in his blog, he said that performance in supercomputers has scaled thanks to new programming models and hardware such as accelerators and graphics cards.

“Code needs to be structured according to concurrency levels, such as in programming languages like the one provided by Barcelona Supercomputing Center’s OmpSS,” Lippert said.

Despite the title of the presentation, the aim is not to challenge Amdahl’s law, Lippert said.

“On the contrary, I think, we are not taking Amdahl’s law seriously enough. It is simply obvious that we should adapt the right piece of hardware to the corresponding concurrency,” Lippert said.

“Only this approach has the potential to be most energy efficient and performance oriented at the same time.”