Kahn Process Networks and a Reactive Extension - Signal Processing Systems - page 1050

Notice that the game as a whole has different modes of streaming (3D graphics,
video), but similarly, components in the stream processing part have different
modes or states of streaming execution. The object renderer for instance can be
reconfigured to different modes depending on the number of objects that need to
be rendered (using the event channel nrObj controlled by the scene graph process).
This illustrates the need for a hierarchical, compositional approach to combining
state/event based models with streaming and data-flow based models, as realized by
the RPN model.

8.2 Design Considerations of RPN

We discuss the main concepts that have had an impact on the design of the model of
Reactive Process Networks.

8.2.1 Streams, Events and Time

Streaming applications represent functions or data transformations. They absorb
input and they produce the corresponding output. There is often no inherent notion
of time, except for the ordering of tokens in the individual data streams. (We are
aiming for an untimed model similar to KPN.) Tokens in different streams have
no relation in time, except for causal relationships implicitly defined by the way
processes operate on the tokens. These process networks are ideally determinate,
i.e., the order and time in which processes execute is irrelevant for the functional
result. The output of a process network is completely determined as soon as the
input is known. The actual computation of this output introduces a certain latency
in the reaction. This latency is not part of the functional specification, but merely a
consequence of the computation process. It may be subject to constraints, such as a
maximum latency. Time is sometimes implicitly present in the intention of steams.
A stream may carry for instance, a sequence of samples of an audio signal that are
1/44100th of a second apart, or video frames of which there are 25 or 30 in every
second. Such streams are called periodic . For final realizations, time-related notions
such as throughput, latency and jitter are of course important.
Events, have a somewhat different relationship to time. An event is unpredictable
and the moment when it arrives, in relation to the streams, is significant, but
unknown in advance. In many event-based models, the synchrony hypothesis
applies, which states that the response to an event can be completed before the
following event arrives or is taken into account. This simplifies specifying how
a system responds to events. A classical model for event-based systems are state
machines, where events make it change from one state into another. Prominent
characteristics are non-determinism and a total ordering of events. If events come
from outside and are not predictable a priori, then the system evolves in a non-
deterministic way under the influence of the events, even if the response to a
particular event is deterministic.