Distributed systems that span large geographic distances or manage large numbers of objects are already common place. In such systems, programming applications with even modest reliability requirements to run correctly and efficiently is a difficult task due to asynchrony and the possibility of complex failure scenarios. We describe the architecture of the RELACS communication subsystem that constitutes the microkernel of a layered approach to reliable computing in large-scale distributed systems. RELACS is designed to be highly portable and implements a very small number of abstractions and primitives that should be sufficient for building a variety of interesting higher-level paradigms.