Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.

Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.

Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.

Author ORCID Identifier



Open Access Dissertation

Document Type


Degree Name

Doctor of Philosophy (PhD)

Degree Program

Computer Science

Year Degree Awarded


Month Degree Awarded


First Advisor

Yanlei Diao

Subject Categories

Computer Sciences | Databases and Information Systems | Systems Architecture


Today, the ability to process "big data" has become crucial to the information needs of many enterprise businesses, scientific applications, and governments. Recently, there have been increasing needs of processing data that is not only "big" but also "fast". Here "fast data" refers to high-speed real-time and near real-time data streams, such as Twitter feeds, search query streams, click streams, impressions, and system logs. To handle both historical data and real-time data, many companies have to maintain multiple systems. However, recent real-world case studies show that maintaining multiple systems cause not only code duplication, but also intensive manual work to partition the analytics workloads and determine which data is processed by which system. These issues point to the need for a general, unified data processing framework to support analytical queries with different latency requirements.

This thesis takes a further step towards building a general, unified system for big and fast data analytics. In order to build such a system, I propose to build on existing solutions on data parallelism and extend them with two new features: incremental processing and stream processing with latency constraints. This thesis starts with Hadoop, the most popular open-source MapReduce implementation, which provides proven scalability based on data parallelism. I answer the following questions: (1) Is Hadoop able to support incremental processing? (2) What are the necessary architecture changes in order to support incremental processing? (3) What are the additional design features needed to support stream processing with latency constraints? The thesis includes three parts that answer each of the questions.

The first part of the thesis validates whether the existing MapReduce implementations can support incremental processing. Incremental processing means that computation is performed as soon as the relevant data becomes available. My extensive benchmark study of Hadoop-based MapReduce systems shows that the widely-used sort-merge implementation for partitioning and parallel processing poses a fundamental barrier to incremental computation. I further propose a cost model, and optimize the Hadoop system configuration based on the model. The benchmark results over the optimized system verify that the barrier to incremental computation is intrinsic, and cannot be removed by tuning system parameters.

In the second part of the thesis, I employ various purely hash-based techniques to enable fast in-memory incremental processing in MapReduce, and frequent key based techniques to extend such processing to workloads that require memory more than available. I evaluate my Hadoop-based prototype equipped with all proposed techniques. The results show that the hash techniques allow the reduce progress to keep up with the map progress with up to 3 orders of magnitude reduction of internal disk spills, and enable results to be returned early.

The third part of the thesis aims to support stream processing with latency constraints based on the incremental processing platform resulted from the second part. I perform a benchmark study to understand the sources of latency. I then propose a number of necessary architecture changes to support stream processing, and augment the platform with new latency-aware model-driven resource planning and latency-aware runtime scheduling techniques to meet user-specified latency constraints while maximizing throughput. Experiments using real-world workloads show that the techniques reduce the latency from tens or hundreds of seconds to sub-second, with 2x-5x increase in throughput. The new platform offers 1-2 orders of magnitude improvements over Storm, a commercial-grade distributed stream system, and Spark Streaming, a state-of-the-art academic prototype, when considering both latency and throughput.