Prof. Daniela Tuninetti’s talk at Motorola

Distributed Coded Computations

Prof. Daniela Tuninetti will give a talk at Motorola on June 28, 2019.

Title: Distributed Coded Computations

Abstract: Since Shannon’s landmark paper, codes have been successfully employed at the physical layer of communications systems to detect or correct errors introduced by various channel impairments, such as random noise. In modern large-scale computing systems, the main problem is however not “noise” but the “distributed nature” of the data availability or of data processors. If “lack” of data at a node in a computing system is thought of as an “erasure,” then codes can come to the rescue. Erasure and network codes are expected to become critical components in distributed storage systems, in distributed computation systems, and in cache-aided systems. In those examples, codes effectively mitigate the effect of strugglers or reduce the communication loads in the data shuffling phase of distributed machine learning algorithms. They do so by carefully leveraging local storage capabilities/redundancy to create multicast opportunities in the communication data exchange phase (i.e., multicast means that the same transmission benefits simultaneously many workers). This talk will focus on the use of codes for speeding up computations in large distributed computing systems, and discuss interesting tradeoffs between distributed storage/cache space and download/delivery time.

This event is part of the activities of the IEEE Chicago Section, Information Theory Chapter