Straggler Mitigation in Distributed Computing

Abstract

While distributed computing is widely implemented in large-scale systems for computation, the straggler effect has become a critical issue to computational performance. It is agreed that the task computation time is dominated by the slowest worker, known as the straggler. The existence of stragglers in a distributed network results in a longer execution time of the tasks and poses a limitation towards the performance. Recent works have proposed several coded distributed computing approaches to alleviate this effect. This study intends to implement straggler mitigation with erasure codes and evaluate the performance of such a strategy in a simulated distributed environment. The proposed system architecture adopts a master-worker model. The master handles the requests and performs data preprocessing and collection while the workers compute the received tasks in parallel. Experiments are conducted by distributing a computational task, such as matrix multiplication, to verify the limitation and effectiveness of the coding strategy. In this report, it is shown that using an (n, k) linear code is effective to mitigate the straggler effect provided that the number of stragglers is smaller than or equal to n – k, with an additional overhead generated by the code which is roughly 8% of the total duration.