ITR: A Grid for Research and Education in Distributed Systems and Networks

University Of Utah, Salt Lake City UT

Investigators

Abstract

Over the last two years, we've built a new type of experimental environment for research in distributed systems and networks: a highly configurable Internet emulator, known as Emulab. This time-and spaceshared facility has been used by numerous top institutions for research published in top venues. So successful is the project, that others are building their own Emulabs, centered around our software and design. While the individual power of such Emulabs in experimentation and teaching is great, connecting these into one large Research and Education Grid will offer unprecedented power and flexibility. We now propose, with dramatically expanded software and innovative algorithms, to establish the framework for a federated set of local Emulabs, each heterogeneous under local control and offering a unique set of hardware. Adding sites to the Grid will be easy and cheap, offering even small institutions a chance to contribute while gaining access to an immensely diverse and powerful resource. Individual sites' users will have priority for local hardware, as well as control over local administrative, security, and resource allocation policies. Such autonomy will encourage organic growth, while intelligent allocation, scheduling, and swapping systems will provide abstraction. Researchers and students can be unconcerned with local complexities, instead seeing one large collection of a wide variety of hardware, including standard PCs, network processors, wireless nodes, and more. Even the links between Emulabs will be a useful resource for experiments wishing to operate on the real Internet, while controlled-bandwidth links will shield experiments that do not. Support for wireless, mobile nodes will open entirely new avenues for experimentation. Achieving this vision, however, involves daunting challenges. Creating the software to manage this will involve a dramatic extension of existing Emulab systems and the application of novel algorithms and techniques. Existing database systems will be augmented with failure-resistant peer-to-peer sharing of network information between Emulabs. Through these mechanisms, topology data will be shared for observation and experiment scheduling. Such scheduling on a non-static, wide-area system is a particularly challenging NP-complete problem which will require new algorithms to be designed and implemented. Automated systems to control experiment scheduling and allocation will be developed all the while facing the challenge of maintaining simple user interfaces. Emulab algorithms and software must also be extended to handle a wider variety of hardware, including wireless nodes. Aspect-oriented programming and component technology, developed at Utah, will manage software interfaces for complex and numerous inter-operating hardware systems. In addition, methods for sharing, saving, and restoring state between machines, as well as maintaining performance isolation while multiplexing multiple virtual nodes on a single computer, will be explored. The incorporation of multiple Emulabs into a single federated entity offers a rich opportunity to overcome a wide variety of network research and software engineering challenges, and allows for experimentation on an unprecedented scale. Users, ranging from individual students to large research groups, will benefit from a greatly expanded ability to controllably and faithfully emulate large real networks, coupled with the availability of diverse and cutting-edge network components. Unifying resources in a standard and consistent way will simplify experimentation, and make accurate emulation an integral part of systems and network instruction and research. The resulting loosely-coupled distributed system will enable experiments in a host of areas. As just one example, if Emulab is successful in one of its core goals, isolation, experiments too dangerous to conduct in the wild could be run. Realistic Internet War Games could be staged, in which attackers and defenders engage in a simulation of possible Internet attack scenarios, replete with destructive and contagious worms and dangerous automated counterattacks. Only in an Emulab Petri dish could they be safely evaluated at reasonable scale.