Author's School

School of Engineering & Applied Science

Author's Department/Program

Computer Science and Engineering

Language

English (en)

Date of Award

Summer 9-1-2014

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Chair and Committee

Raj Jain

Abstract

In this thesis we present the architecture, design, and prototype implementation details of AppFabric. AppFabric is a next generation application delivery platform for easily creating, managing and controlling massively distributed and very dynamic application deployments that may span multiple datacenters.

Over the last few years, the need for more flexibility, finer control, and automatic management of large (and messy) datacenters has stimulated technologies for virtualizing the infrastructure components and placing them under software-based management and control; generically called "Software-defined Infrastructure" (SDI). However, current applications are not designed to leverage this dynamism and flexibility offered by SDI and they mostly depend on a mix of different techniques including manual configuration, specialized appliances (middleboxes), and (mostly) proprietary middleware solutions together with a team of extremely conscientious and talented system engineers to get their applications deployed and running. AppFabric, 1) automates the whole control and management stack of application deployment and delivery, 2) allows application architects to define logical workflows consisting of application servers, message-level middleboxes, packet-level middleboxes and network services (both, local and wide-area) composed over application-level routing policies, and 3) provides the abstraction of an "application cloud" that allows the application to dynamically (and automatically) expand and shrink its distributed footprint across multiple geographically distributed datacenters operated by different cloud providers. The architecture consists of a hierarchical control plane system called "Lighthouse" and a fully distributed data plane design (with no special hardware components such as service orchestrators, load balancers, message brokers, etc.) called "OpenADN". The current implementation (under active development) consists of ~10000 lines of python and C code.

AppFabric will allow applications to fully leverage the opportunities provided by modern virtualized Software-Defined Infrastructures. It will serve as the platform for deploying massively distributed, and extremely dynamic next generation application use-cases, including:

Internet-of-Things/Cyber-Physical Systems: Through support for managing distributed gather-aggregate topologies common to most Internet-of-Things(IoT) and Cyber-Physical Systems(CPS) use-cases. By their very nature, IoT and CPS use cases are massively distributed and have different levels of computation and storage requirements at different locations. Also, they have variable latency requirements for their different distributed sites. Some services, such as device controllers, in an Iot/CPS application workflow may need to gather, process and forward data under near-real time constraints and hence need to be as close to the device as possible. Other services may need more computation to process aggregated data to drive long term business intelligence functions. AppFabric has been designed to provide support for such very dynamic, highly diversified and massively distributed application use-cases.

Network Function Virtualization: Through support for heterogeneous workflows, application-aware networking, and network-aware application deployments, AppFabric will enable new partnerships between Application Service Providers (ASPs) and Network Service Providers (NSPs). An application workflow in AppFabric may comprise of application services, packet and message-level middleboxes, and network transport services chained together over an application-level routing substrate. The Application-level routing substrate allows policy-based service chaining where the application may specify policies for routing their application traffic over different services based on application-level content or context.

Virtual worlds/multiplayer games: Through support for creating, managing and controlling dynamic and distributed application clouds needed by these applications. AppFabric allows the application to easily specify policies to dynamically grow and shrink the application's footprint over different geographical sites, on-demand.

Mobile Apps: Through support for extremely diversified and very dynamic application contexts typical of such applications. Also, AppFabric provides support for automatically managing massively distributed service deployment and controlling application traffic based on application-level policies. This allows mobile applications to provide the best Quality-of-Experience to its users without

This thesis is the first to handle and provide a complete solution for such a complex and relevant architectural problem that is expected to touch each of our lives by enabling exciting new application use-cases that are not possible today. Also, AppFabric is a non-proprietary platform that is expected to spawn lots of innovations both in the design of the platform itself and the features it provides to applications. AppFabric still needs many iterations, both in terms of design and implementation maturity. This thesis is not the end of journey for AppFabric but rather just the beginning.

DOI

https://doi.org/10.7936/K7CJ8BJH

Comments

Permanent URL: http://dx.doi.org/10.7936/K7CJ8BJH

Share

COinS