1. com-pos-ite (noun): made up of disparate or separate parts or elements; compound
2. compose-it (command): assembling a compound from discrete elements or components

The Composite system is motivated by the desire to construct the operating system from fine-grained components. Thus, all policies within the system are defined as services at user-level. Composite includes novel mechanisms for dynamically creating and removing protection domain boundaries in the system in response to application performance goals. Additionally, Composite enables the user-level definition of a system scheduling policies, thus allowing them to be customized, like all other policies in the system.

Using careful interposition techniques, many of the policies within a host kernel can be hijacked and redefined at user-level. This allows application-specific services to be defined in normal OSes such as Linux.

User-Level Sandboxing (ULS) involves providing an execution environment in a COTS system (Linux on x86) that is isolated from the core kernel, but can execute in a first-class manner. Applications with demanding temporal constraints that cannot be met using heavy-weight processes can execute portions of their code in the sandbox. Invocation of these services is first-class as they can be executed directly by the base kernel, while maintaining isolation of the system from the extensions. Code deployed in a Sandbox can be invoked while in any process's context. This ability can be used to, at interrupt/bottom half time, modify CPU allocations, or to process network data. The group has published twice demonstrating that low-latency processing can be obtained using ULS, both in a Real-time environment, and in a customizable networking stack environment.

Work done in Spring '05 on simulating different multicast tree construction methods at the application level. The interesting part of this work is the tomography work necessary to configure the multicast tree on top of the actual network infrastructure in a manner that preserves many of the characteristics of the physical network. Application-level multicast's main concern is to create a good mapping between the logical tree and the underlying network while trading off data delivery latency, link stress, and tree construction time. We investigate multiple methods and demonstrate via simulation which excel at which dimensions. Many of the methods are on par with, or outperform many competing multicast tree construction methods.