Hierarchical Coded Matrix Multiplication

In distributed computing systems slow working nodes, known as stragglers, can greatly extend finishing times. Coded is a technique that enables straggler-resistant computation. Most coded techniques presented to date provide robustness by ensuring the time finish depends only on set of fastest nodes. However, while stragglers do compute less work than non-stragglers, in real-world commercial cloud (e.g., Amazon’s Elastic Compute Cloud (EC2)) distinction often soft one. this paper, we develop hierarchical exploits completed all both fast and slow, automatically integrating potential contribution each. We first present conceptual framework represent division amongst nodes matrix multiplication cuboid partitioning problem. This allows us unify existing methods motivates new techniques. then three hierarchical term xmlns:xlink="http://www.w3.org/1999/xlink">bit-interleaved computation (BICC), xmlns:xlink="http://www.w3.org/1999/xlink">multilevel (MLCC), xmlns:xlink="http://www.w3.org/1999/xlink">hybrid (HHCC). paradigm, each worker tasked with completing sequence (a hierarchy) ordered subtasks. The subtasks, complexity each, designed so partial be used, rather ignored. note our used conjunction any method. illustrate showing how use accelerate previously developed enabling them exploit stragglers. Under widely studied statistical model completion time, approach realizes 66% improvement expected time. On Amazon EC2, gain was 27% when are simulated.