Straggler Mitigation through Unequal Error Protection for Distributed Matrix Multiplication
share
Large-scale machine learning and data mining methods routinely distribute computations across multiple agents to parallelize processing. The time required for computation at the agents is affected by the availability of local resources: this gives rise to the "straggler problem" in which the computation results are held back by unresponsive agents. For this problem, linear coding of the matrix sub-blocks can be used to introduce resilience toward straggling. The Parameter Server (PS) codes the matrix products and distributes the matrix to the workers to perform multiplication. At a given deadline, it then produces an approximation the desired matrix multiplication using the results of the computation received so far. In this paper, we propose a novel coding strategy for the straggler problem which relies on Unequal Error Protection (UEP) codes. The resiliency level of each sub-block is chosen according to its norm, since blocks with larger norms affect more the approximation of the matrix multiplication. We validate the effectiveness of our scheme both theoretically, as well as through numerical evaluations. We derive a theoretical characterization of the performance of UPE using random linear codes and compare it the case of equal error protection. We also apply the proposed coding strategy to the computation of the back-propagation step in the training a Deep Neural Network (DNN). In this scenario, we investigate the fundamental trade-off between precision of the updates versus time required for their computation.
READ FULL TEXT