Leveraging Coded Computing with Compact Data Structures to Mitigate Straggler Problems in Distributed Edge Learning

Implementing distributed edge learning is challenging and can frequently face straggler phenomena caused not only by heterogeneous compute capabilities but also by unstable network conditions and constrained communication resources. Although coded computing techniques have been widely studied for straggler mitigation, existing approaches mainly resort to computation redundancy and largely overlook communication overhead in edge environments. This paper proposes a new direction that combines coded computing with compact data structure (CDS) techniques to mitigate straggler problems in distributed edge learning. To validate this novel idea, we are conducting a prototyping study on a Raspberry Pi edge cluster to evaluate representative coded computing methods together with the CDS technique Count-Min Sketch on distributed ResNet-18 training over CIFAR-10. Although the development of a complete solution prototype is still in progress, our current experimental results have shown that CDS-based gradient compaction can significantly reduce communication overhead and training latency while maintaining acceptable model accuracy. To our best knowledge, this is the first work to explore combining coded computing with CDS for straggler mitigation in distributed edge learning.

Recommended citation: Aodhan Ferry, Shishir Nagaraja, Yun Wu, Rajiv Ranjan, Zheng Li. (2026). "Leveraging Coded Computing with Compact Data Structures to Mitigate Straggler Problems in Distributed Edge Learning." International Conference on Algorithms and Architectures for Parallel Processing.
Download Paper | Download Slides