Professor Keijo Heljanko: Effective computing requires cooperation between computers

19.11.2018
Increasingly demanding calculations require distributed computing, or several computers focused on solving the same problem together. However, computers cannot agree on the distribution of duties without help from data scientists.

Big problems require powerful computers. If a certain problem is too extensive for even a powerful individual computer, more than one of such machines will be needed.

This sounds like something that should be taken for granted, yet the truth is actually something else, according to Keijo Heljanko, a new professor of parallel and distributed data science. Heljanko is also serving as the deputy director of HiData, the Helsinki Centre for Data Science, a unit jointly coordinated by the University of Helsinki and Aalto University.

The way in which computers share information and have a dialogue when they carry out computing greatly influences its effectiveness.

“Parallel computing, where several computers process the same problem, is not easy to organise. The way in which computers share information and have a dialogue when they carry out computing greatly influences its effectiveness.”

This is an issue on which Heljanko has worked since the turn of the millennium. At first, he analysed systems of parallel computing, developing appropriate tools, but later he moved on to develop distributed computing systems himself.

The problem is not only academic, for two distinct reasons. The first is the continuously growing masses of data. For example, Google searches and the analysis of genetic data already take up so much computing power that individual computers are not up to the task. According to estimates, the amount of data will globally grow tenfold between 2013 and 2020, and there is no end in sight. New features require more data.

“Take automated face recognition, which could be applied, for example, to data security in various ways. Introducing facial recognition features to everyday devices would, however, require vast amounts of computing power.”

Devices heating up is the other reason. Earlier, increasing demand for computing power could be sated by accelerating the clock speed of microprocessors. In practice, this translates into cramming more and more transistors into the same space. However, today’s processors hold so many transistors that soon their numbers can no longer be increased.

Therefore, the demand for increasingly greater computing power threatens to result in overheating computers.

A solution is provided by computers optimised for certain tasks, such as developing artificial intelligence software. The processing power of computers designed for such limited purpose can still be advanced further.

However, using optimised computers requires linking computers of different types together in such a manner that each machine only focuses on the task where its particular strength lies. This, in turn, requires better methods of parallel computing.

“The significance of parallel computing will increase in the future, since there really are no alternatives for processing constantly and rapidly growing data masses,” says Heljanko.

More about the subject: Data science news