Distributed Computing Under Communication Constraints (IMAGE)
Caption
A new study by Botond Szabo (Bocconi Department of Decision Sciences) lays the cornerstone for more accurate, reliable, and interpretable distributed computing methods.
In the world of big data, when the need arises to estimate many parameters in very complex statistical models which make use of large amounts of available information, computation time becomes unsustainable even with the fastest supercomputers. One of the strategies developed to cope with the issue is distributed (or parallel) computing. Data (or tasks, in some cases) are divided among many machines and only summary information (the results of computations) is sent to a central location, say a meteorological station, an astronomy observatory, or a traffic control system. This method also mitigates privacy concerns since most data don’t have to be moved around.
In any case, even communicating only summary information between servers can be costly, sostatisticians have borrowed from electric engineers the idea of bandwidth limitation. “The goal,” says Professor Szabo, “is to minimize the flow of data, losing as little information as possible. Furthermore, parallel computing is often a black-box procedure, i.e. a procedure which transforms inputs into outputs in not-well-understood ways, and this makes results neither completely interpretable, nor reliable. Finding mathematical models which give theoretical underpinnings to such procedures would be desirable.” Prof. Szabo is the recipient of a ERC Grant aimed at tackling such issues.
In his paper with Lasse Vuursteen (Delft University of Technology) and Harry van Zanten (Vrije Universiteit Amsterdam), Prof. Szabo derives the best tests to minimize the loss of information in a distributed framework where the data is split over multiple machines and their communication to a central machine is limited to a given quantity of bits.
To know more: https://www.knowledge.unibocconi.eu/notizia.php?idArt=25717
Credit
Weiwei Chen, Bocconi University
Usage Restrictions
For journalistic purposes only
License
Original content