`

Timezone: »

 
Poster
Communication trade-offs for Local-SGD with large step size
Aymeric Dieuleveut · Kumar Kshitij Patel

Wed Dec 11 05:00 PM -- 07:00 PM (PST) @ East Exhibition Hall B + C #155
Synchronous mini-batch SGD is state-of-the-art for large-scale distributed machine learning. However, in practice, its convergence is bottlenecked by slow communication rounds between worker nodes. A natural solution to reduce communication is to use the \emph{``local-SGD''} model in which the workers train their model independently and synchronize every once in a while. This algorithm improves the computation-communication trade-off but its convergence is not understood very well. We propose a non-asymptotic error analysis, which enables comparison to \emph{one-shot averaging} i.e., a single communication round among independent workers, and \emph{mini-batch averaging} i.e., communicating at every step. We also provide adaptive lower bounds on the communication frequency for large step-sizes ($ t^{-\alpha} $, $ \alpha\in (1/2 , 1 ) $) and show that \emph{Local-SGD} reduces communication by a factor of $O\Big(\frac{\sqrt{T}}{P^{3/2}}\Big)$, with $T$ the total number of gradients and $P$ machines.

Author Information

Aymeric Dieuleveut (Ecole Polytechnique, IPParis)
Kumar Kshitij Patel (Toyota Technological Institute at Chicago)

More from the Same Authors