Separating Presentation from Interface in RPC and IDLs

In RPC-based communication, we term the interface the set of remote procedures and the types of their arguments; the presentation is the way these procedures and types are mapped to the target language environment in a particular client or server, including semantic requirements. For example, presentation includes the local names assigned to RPC stubs, the physical representation of a logical block of data (e.g., in-line, out-of-line, linked blocks), and trust requirements (e.g., integrity, security). In existing systems, the presentation of a given RPC construct is largely fixed. Separating presentation from interface, both in the interface definition language (IDL) itself and in the RPC implementation, is the key to interoperability, with many benefits in the area of elegance, as well. This separation and resulting cleanliness makes it manageable to generate specialized kernel code paths for each type of clientserver pair. This is a key element of end-to-end optimization. The separation should also allow the integration of disparate RPC optimization techniques, such as those applied in LRPC[2] and fbufs[6], into a single system, in a uniform and fully interoperable way. In initial work we demonstrate a variant of threaded code generation and two presentation-based optimizations, transparently activated by the RPC system. Each of these optimizations speeds up local RPC by approximately 25%.1