Introduction

The emergence of new networking technologies such as broadband to the home, Wireless 3G [18], and Bluetooth [10], coupled with increasing numbers of network-capable end devices holds the potential for future application services that significantly enhance user experience by providing seamless, ubiquitous access. To take an example, consider the following scenario. Alice, a telecommuting employee, starts her day by initiating a teleconference on her laptop connected to the internet using a wired LAN. During the conference, a hub failure renders the wired LAN unavailable. Fortunately, the service detects this, and seamlessly switches data transmission to a local wireless network while simultaneously degrading picture quality upon recognizing that the wireless LAN has insufficient bandwidth for continuous video at the original resolution and rate. Shortly after, Alice leaves her office to meet a client. She shuts down her laptop, and resumes the teleconference in her car using a PDA connected to a metro-area wireless network. The service further downgrades the media stream (say to only include audio), while recording the full stream at a server that Alice can check offline. Although the above scenario is compelling, its requirements -- rapid creation and deployment of new services, application-aware computation in the network, and dynamic and distributed adaptation -- are poorly handled by current internet infrastructure. Moreover, the existing view which hides network characteristics from the application and treats services as standalone entities is incompatible with the large variation in network and end-device characteristics. Current-day data paths can include links with very different bandwidth, delay, and error characteristics, ranging from serial links to wireless to broadband to fiber. Hiding these differences from the application will result in unsatisfactory application performance, and the alternative of providing differentiated service for different networks/end-devices cannot adequately cope with dynamically changing environments. One solution to these problems is to inject additional functionality into the network that can dynamically adapt to resource characteristics of end-devices and network links by handling activities such as protocol conversion, data transcoding, etc. Several researchers have proposed infrastructures for achieving this goal, ranging from end-point solutions [12,15] to more distributed alternatives that introduce application-aware functionality either at the network level [17,3,20] or at the application level [1,6,8]. Although these systems have articulated a common set of high-level architectural requirements, many challenges, particularly with respect to dynamic services management and composition, remain before the infrastructures see widespread deployment. This paper describes Composable Adaptive Network Services (CANS), an application-level infrastructure for customizing the data path between client applications and services, which focuses on three such challenges:

• Efficient and Dynamic Composition, enabling separately defined components to be dynamically instantiated and interconnected using efficient mechanisms (e.g., shared memory within a host).
• Dynamic and Distributed Adaptation, enabling adaptation to environment changes along the entire data path while incurring low overhead and maintaining overall application semantics.
• Support for Legacy Applications and Services, enabling the latter to be integrated into CANS with minimal effort. Requiring rewrites of each application and service is neither practical nor desirable.
  

CANS addresses these challenges by constructing networks that include applications, stateful services, and data paths between them built up from mobile soft-state objects called drivers. Drivers implement a standard interface, permitting efficient composition and semantics-preserving adaptation. Both services and data paths can be dynamically created and reconfigured: a planning and event propagation facility enables distributed adaptation, and a flexible type-based composition model dictates how new services and drivers are integrated with existing ones. CANS provides three adaptation modes to permit cost-functionality tradeoffs: intra-component, by reconfiguring data paths, and by creating new services and data paths. Legacy components plug into CANS using delegation and an interception layer that transparently virtualizes network bindings, currently TCP sockets. CANS has been implemented on Windows 2000 clients and Java/RMI-capable intermediate hosts. Each node runs the CANS execution environment, which supports dynamic creation, migration, and adaptation of drivers and services. Experience with a case study involving a shrink-wrapped application (Windows MediaPlayer) in a dynamically changing network environment indicates the potential of our approach: CANS permits dynamic deployment and distributed adaptation of application-aware components to improve overall user experience. The rest of this paper is organized as follows. Section 2 presents the CANS architecture, with details about its components and distributed adaptation support appearing in Sections 3 and 4. Section 5 presents the CANS implementation and the MediaPlayer case study. Section 6 discusses related efforts, and Section 7 concludes.