Dynamic re-optimization techniques for stream processing engines and object stores

Naresh Kumar Reddy Rapolu, Purdue University

Abstract

Large scale data storage and processing systems are strongly motivated by the need to store and analyze massive datasets. The complexity of a large class of these systems is rooted in their distributed nature, extreme scale, need for real-time response, and streaming nature. The use of these systems on multi-tenant, cloud environments with potential resource interference necessitates fine-grained monitoring and control. In this dissertation, we present efficient, dynamic techniques for re-optimizing stream-processing systems and transactional object-storage systems. In the context of stream-processing systems, we present VAYU, a per-topology controller. VAYU uses novel methods and protocols for dynamic, network-aware tuple-routing in the dataflow. We show that the feedback-driven controller in VAYU helps achieve high pipeline throughput over long execution periods, as it dynamically detects and diagnoses any pipeline-bottlenecks. We present novel heuristics to optimize overlays for group communication operations in the streaming model. In the context of object-storage systems, we present M-Lock, a novel lock-localization service for distributed transaction protocols on scale-out object stores to increase transaction throughput. Lock localization refers to dynamic migration and partitioning of locks across nodes in the scale-out store to reduce cross-partition acquisition of locks. The service leverages the observed object-access patterns to achieve lock-clustering and deliver high performance. We also present TransMR, a framework that uses distributed, transactional object stores to orchestrate and execute asynchronous components in amorphous data-parallel applications on scale-out architectures.

Degree

Ph.D.

Advisors

Grama, Purdue University.

Subject Area

Computer science

COinS