The Impact of Inexpensive Communication on a Commercial RPC System

Distributed object systems and the underlying remote procedure call layers which implement them have been engineered for expensive communication (high overhead, low-speed networks with unreliable delivery). The architecture of these RPC systems embody design choices that incur communication overhead , obviating the beneets of inexpensive communication (low overhead, low latency, high bandwidth). Increased use of RPC in environments where communication is inexpensive (e.g. clusters and high speed LAN's connected by VIA 20] networks) provides the impetus for reengineering RPC systems to deliver the beneets of inexpensive communication. To explore the implications of inexpensive communication, we study an implementation MSRPC atop Illinois Fast Messages 14], a high performance, user-level networking kernel. We integrate FM into the MSRPC system by building a datagram loadable transport, and while the latency of FM is 11 s and the loadable transport 15 s, insertion of the fast network only reduces the null RPC time by three times, from 362 to 105 s. Detailed prooling shows that existing RPC systems are poorly designed to exploit inexpensive communication , implementing the key design choices of transport management, endpoint/name management, thread management, and call management under the assumptions of unreliable networks, shared end-points, hiding communication latency with context-switching, and complex error situations. Speciically, our prooling studies attributed 60 of overhead per call to thread management and 28 s per call to call management. To deliver underlying communication performance to RPC applications, future RPC architectures must be engineered for low latency and exploit System Area Network reliability and simpler error conditions to simplify protocols.