In this paper we discussed the design and implementation of an elastic management and control architecture (MCA) called Sandman which enables applications to inject their own policies into the network in the form of DLAs. It allows very fine grain control over the resources in the network (up to the level of individual switch connections) and allows extensions by means of new user-defined operations accessible to any application. Also, here as elsewhere, we have maintained a strict separation between control path and data path. The resources in the network can be recursively (re-)partitioned at an almost arbitrarily fine granularity using netlets. A netlet can be associated with application-specific code, if this is so desired. A general solution for interoperability across multiple domains has been presented. The solution consists of a simple hop-by-hop mechanism in addition to dedicated signalling channels between Sandman islands (separated by other MCAs) that does not suffer from functionality degradation at the domain boundaries. Finally, we have shown how we can achieve application-specific policies spanning multiple domains by allowing dynamically loadable agents migrate from one MCA to the next, installing application-specific policy in each domain.
We believe that such an open approach to network control and management allows applications to exploit application-specific knowledge that would be difficult to capture by a fixed set of high-level primitive operations. This in turn allows for a class of operations that hitherto have been impossible to implement.