Coded Distributed Computing With Partial Recovery

Coded computation techniques provide robustness against straggling workers in distributed computing. However, most of the existing schemes require exact provisioning straggling behavior and ignore computations carried out by workers. Moreover, these are typically designed to recover desired results accurately, while many machine learning iterative optimization algorithms, faster approximate solutions known result an improvement overall convergence time. In this paper, we first introduce a novel coded matrix-vector multiplication scheme, called xmlns:xlink="http://www.w3.org/1999/xlink">coded with partial recovery (CCPR) , which benefits from advantages both uncoded schemes, reduces time decoding complexity allowing trade-off between accuracy speed computation. We then extend approach implementation more general tasks proposing communication scheme recovery, where subtasks computed before being communicated. Numerical simulations on large linear regression task confirm proposed terms latency.