System and method for managing distribution of computations in multi-antenna and multi-transmitter environments

Elia, Petros; Lampiris, Eleftherios; Parrinello, Emanuele
Patent N°WO/2019/110599, 13 June 2019

In the field of distributed computing, the communication that needs to be performed between the computing nodes or helper nodes, poses a major performance bottleneck. In order to reduce the time associated with this so-called shuffling phase of inter-node communication, it was recently proposed that the computational redundancy across the nodes be exploited to allow – via a specific type of coded communications – that each transmission becomes useful to more than one node at a time, thus significantly reducing the overall communication time. But the authors in this patent application here, have identified as a main bottleneck of such coded' current approaches, the fact that the input dataset to be processed for a given job, has to be split into an exponentially large number of smaller subsets, which is often impossible, thus significantly constraining the speedup gains attributed to coded communication. The inventors here have invented a new method of assigning data to the computing nodes, which dramatically alleviates the aforementioned bottleneck, thus resulting in dramatic reduction to the overall execution time of the distributed computing job.


Type:
Brevet
Date:
2019-06-13
Department:
Systèmes de Communication
Eurecom Ref:
6091
Copyright:
© EURECOM. Personal use of this material is permitted. The definitive version of this paper was published in Patent N°WO/2019/110599, 13 June 2019 and is available at :

PERMALINK : https://www.eurecom.fr/publication/6091