New Internet Computing Lab (NICL)
NICL explores system software infrastructures for the New Internet. We focus on hosting platforms for network services, including network computing and network storage. In particular, our research deals with networked utility systems that share and provision a virtualizable resource substrate (computers, network elements, storage) on demand for a wide range of networked services. Hosted services (guests) can include computational grids and cloud computing middleware and also new kinds of networked systems that run the substrate resources under their control in entirely different ways. We proposed the notion of computing on a "server cloud" in talks in 2001-2002 (Servers in the Mist [PDF]) and it has been a cornerstone of our research agenda ever since.A key research challenge is to build autonomic services and cloud control systems that can manage themselves and hold human administrative burdens constant as the system scales. The real promise of networked cloud computing is to enable dynamic, adaptive network application services that can deploy wherever resources are available and demand exists, and can sense-and-respond to adapt automatically to changes in traffic demands or resource conditions.
If we are successful, utility substrates and control services will evolve into an open public infrastructure requiring flexible, secure, robust, and decentralized control. The control architecture must resolve the "tussle" of contending demands, changing priorities, and rapidly advancing technology, all within the framework of a self-sustaining system.
Projects
- Orca: Open Resource Control Architecture. Orca is a
control plane architecture for an Internet operating system based on a foundational abstraction of resource
leasing. Orca is also an open-source
software release incorporating the latest software from several earlier and ongoing projects:
- Shirako is a resource leasing toolkit implemented in Java. It defines Java interfaces to incorporate new resources and services into an Orca control plane.
- Cluster-On-Demand (COD) is a site manager for reconfigurable mixed-use clusters. COD guests use Shirako leasing interfaces to allocate and configure virtual clusters.
- Automat is an interactive Web-based laboratory for autonomic services and data centers. It includes a Web portal for deploying and monitoring controller modules that manage dynamic resource provisioning for Shirako/COD sites and the services deployed on them. It also provides a test harness for subjecting services and controllers to various workloads and faultloads. See the paper from Hot Topics in Autonomic Computing (HOTAC 2007) [PDF].
- SHARP is an architecture for secure highly available resource peering. SHARP defined an initial form of the leasing abstraction used in Orca, and secure delegation mechanisms to exchange leases as accountable contracts among self-interested actors.
- Virtual grids and cloud computing middleware. The Orca software is a platform for diverse guest environments that schedule workloads on "clouds" of infrastructure resources. These include job management services, virtual desktop services, and middleware for data-intensive cluster computing, such as Hadoop/MapReduce. We have run several such systems as pluggable "guest packages" above the guest-neutral Orca platform: Globus, Sun GridEngine (SGE), Hadoop, and a VM-based job execution service called JAWS. We are exploring how to configure guests on demand, share resources among multiple guest instances on a common substrate, and adapt resource assignments as conditions change. We also collaborate with the Virtual Workspaces project.
- Statistical learning for performance control, diagnosis, and repair. NIMO (NonInvasive Modeling for Optimization) investigates statistical approaches to learn models of application behavior by observing guest applications under various conditions. These models can serve as a basis for effective and efficient resource management and fault diagnosis in networked utilities. We are also exploring approaches workbench automation that use active learning and proactive design of experiments to learn such models quickly.
- CATS (Certified Authenticated Tamper-Evident Services) investigates fundamental structures for accountable network services in which misbehavior by a participant is detectable by other participants and provable to a third party. Accountability is important for any network service with multiple self-interested actors, including Orca systems with contract-based resource sharing.
- Economic resource exchange. Cereus investigates market-based mechanisms with virtual currencies as a basis for allocating resources in a shared substrate. Such schemes may be implemented as collections of controller modules in an Orca system that supports accountable contracts.
- Green Computing. A key advantage of flexible compute clouds is the ability to manage server resources to save energy. Beginning in 2001, we wrote a series of papers on adaptive thermal/power management for data centers, most in a collaboration with recent PhD Justin Moore (now at Google) and Partha Ranganathan of HP Labs, who was named one of the world's top young innovators in 2007.
People
Faculty and Staff
Collaboratoring faculty at Duke: Staff:- David Becker
Graduate Students
Recently graduated PhD students:- Dr. Laura Grit. Thesis: Extensible Resource Management for Networked Virtual Computing. (Now at Amazon)
- Dr. David Irwin. Thesis: An Operating System Architecture for Networked Server Infrastructure. (Now at U. Mass Amherst)
- Dr. Piyush Shivam. Thesis: Proactive Experiment-Driven Learning for System Management. (Now at Sun Microsystems)
- Dr. Justin Moore. Thesis: Automated Cost-Aware Data Center Management. (Now at Google)
Funding
We are grateful to the National Science Foundation for our major funding. NSF has funded our work most recently through:- CNS-0720829, CSR-VCM: Foundations for a Programmable Self-Managing Hosting Center, with Shivnath Babu.
- CNS-0509408, CSR-AES: Virtual Playgrounds: Making Virtual Distributed Computing Real, with Kate Keahey and Ian Foster.
- ANI 03-30658, NMI: Collaborative Research: A Grid Service for Dynamic Virtual Clusters