The performance of many distributed system services, such
as atomic broadcast, group membership, clock synchro-
nization, etc., depends crucially on the message transmis-
sion delay, either one-way or round-trip, as experienced
by the server processes communicating across the network.
Usually the service specification assumes an a priori given
constant as the maximum acceptable message delay, and
leaves its final determination up to the service implementor.
Often a timeout based on the transmission delay is the only
parameter subject to adjustment, and its value determines
to a great extent the runtime behavior of the service. Be-
cause of such important performance implications, a very
careful choice must be made here, taking many aspects of
the target system in consideration. This paper proposes a
framework for the solution of this problem in networks of
workstations, and illustrates its application by a suite of ex-
periments performed in an environment consisting of Sun
SPARCstation IPX machines, interconnected by an Ether-
net LAN.