• Diese Seite als PDF öffnen.

Gradient coding for straggler mitigation in distributed computing

Description

yle="margin: 0px; padding: 0px; color: #172b4d; font-family: Arial, sans-serif; font-size: 14px;">With the growth of machine learning applications and in the era of big data, the interest towards distributed computing has risen. However, with the splitting of a highly complex computation into lesser ones and computation of those in parallel from external worker nodes, a lot of issues have been encountered. One of them is the presence of straggler nodes, i.e., worker nodes which are slower than the other ones. Therefore, techniques of straggler mitigation have been studied recently. One approach is gradient coding, which is used also in distributed learning.

yle="margin: 10px 0px 0px; padding: 0px; color: #172b4d; font-family: Arial, sans-serif; font-size: 14px;">The tasks of the student is to understand how gradient coding helps in straggler mitigation [1] and how the coding theory tools can help in designing gradient codes [2].

yle="margin: 10px 0px 0px; padding: 0px; color: #172b4d; font-family: Arial, sans-serif; font-size: 14px;"> 

yle="margin: 0px; padding-left: 22px; color: #172b4d; font-family: Arial, sans-serif; font-size: 14px;">
• [1] Tandon, Rashish et al. - Gradient Coding: Avoiding Stragglers in Synchronous Gradient Descent yle="color: #3070b3; text-decoration-line: none; border-bottom: 1px solid gray; padding-bottom: 1px; transition: all 188ms ease 0s;" href="https://arxiv.org/pdf/1612.03301.pdf" rel="noopener" target="_blank">https://arxiv.org/pdf/1612.03301.pdf
• [2] Raviv, Netanel et al. - Gradient Coding from Cyclic MDS Codes and Expander Graphs yle="color: #3070b3; text-decoration-line: none; border-bottom: 1px solid gray; padding-bottom: 1px; transition: all 188ms ease 0s;" href="https://arxiv.org/pdf/1707.03858.pdf" rel="noopener" target="_blank">https://arxiv.org/pdf/1707.03858.pdf

Prerequisites

Contact

Supervisor: