Utility Scheduling for Multi-Tenant Clusters

Abstract

The rapid increase in data size along with the complex patterns of data usage amongst

data scientists presents new challenges for large-scale data analytics systems. Modern dis-

tributed computing frameworks must support complex applications that range from answer-

ing database queries to training machine learning models. As data centers have grown,

managing their resources has become an increasingly important task. New applications

have become popular that make traditional scheduling systems inadequate.

In this thesis, we present distributed scheduling systems aimed at increasing cluster

resource utilization by taking advantage of specific characteristics of data processing ap-

plications. First, we identify a set of applications whose characteristics make them prime

targets for utility-based scheduling. We then focus on two specific types of these applica-

tions in the following systems:

(i) SLAQ: a cluster scheduling system for machine learning (ML) training jobs that

aims to maximize the qualities of all models trained. In exploratory model training, models

can be improved more quickly by redirecting resources to jobs with the highest potential for

improvement. SLAQ reduces latency and maximizes the quality of models being trained

by a distributed ML cluster.

(ii) ReLAQS: a cluster scheduling system for incremental approximate query process-

ing (AQP) systems that aims to minimize the error of all approximate results. In AQP,

queries compute approximate results by sampling data. In AQP, error can be reduced more

quickly by allocating resources to queries with higher error. ReLAQS reduces the latency

required to reach a query result with a given level of error in a shared AQP environment.

These works demonstrate a novel set of methods that can be used in fine-grained

scheduling to build responsive, efficient distributed systems. We have evaluated these

systems on standard benchmark workloads and datasets, as well as popular ML algorithms,

and show both reduced latency and increased accuracy of intermediary results.