Coding and resource allocation for delay-sensitive distributed cloud computing system

Abstract

Coding has been widely used as a way to mitigate the effect of system noises in many engineering applications. Research has also been conducted to distribute large-scale computations required for machine learning and data analytics to the computing clusters such as Amazon EC2. In this dissertation, we show that the latency due to the computation scale and complexity can be dramatically reduced by coding under the distributed cloud computing environment. In the first part, we suggest the optimal load allocation for coded distributed computing with heterogeneous workers. Specifically, we focus on a scenario in which the workers with a same computing capability belong to a group for analysis. We find that load balancing among the groups plays an important role in minimizing the latency. Based on the load balancing, we demonstrate that the distribution of time for each group to complete the given tasks is asymptotically constant. Moreover, we suggest the theoretic lower bound for the expected latency and show that the proposed load allocation asymptotically achieves the lower bound. In the second part, as a generalization of the first part, we propose the optimal worker assignment and load allocation to minimize the expected latency in the presence of multiple masters. Based on the optimal load allocation in the first part, we find the necessary condition for worker assignment to minimize the time required for multiple masters to complete the given tasks. Using the necessary condition, we obtain the optimal worker assignment and load allocation. Furthermore, we show that the optimal worker assignment and load allocation achieve the lower bound of the expected latency, which is shown to be the theoretic limit. In the third part, we consider the problem of estimating the time-dependent ability of workers participating in the distributed computing over heterogeneous clusters. We propose a method for estimating the workers' ability based on the expectation maximization algorithm combined with the particle method. Exploiting the proposed estimation of the workers' ability, one can devise the optimal load allocation and worker assignment for minimizing the latency in the heterogeneous distributed computation with the workers having time-varying abilities. Lastly, we propose code designs to achieve low latency in a communication network environment between a master and a cloud. The proposed codes are designed by applying the interpolation of Reed-Muller and polar code. In addition, the algebraic structures of the proposed codes are analyzed.