A Study of Computational Reconfiguration in a Process Network

In the process network model, the network evolves by reconfiguration. Reconfiguration changes the representation of the network. With a multi-threaded implementation of the process network system, it is necessary to coordinate concurrent accesses to the representational structure. We compare two approaches to the problem of ensuring consistency of representation during reconfiguration. One is a “localized” view, taken from the viewpoint of the process undergoing reconfiguration. The other requires a “global” view of the entire process network structure. We show how reconfiguration takes place in each case, and compare advantages and disadvantages of each. 1 A Technique for Process Network Reconfiguration The Kahn model of process networks [Kah74] is used to represent transformations to streams of data. Process networks are structured as directed graphs where a node represents a process and an edge represents the flow of data from one process to another. Kahn defines a process to be a mapping from one or more input streams (or histories) to one or more output streams. Processes communicate only through directed first in first out (FIFO) stream of tokens with unbounded capacity such that each token is produced exactly once, and consumed exactly once. Production of tokens is non-blocking, while consumption from an empty stream is blocking. The model is highly concurrent and deterministic, and is of interest to us as a semantically sound formulation of flow-based systems. Reconfiguration is a fundamental element of process network semantics as defined by [Kah74], via the recursive process schema. In fact, a process network can be regarded as starting with one node and expanding as the computation proceeds [KM77]. The behaviour of process networks is dynamic, evolving in a top-down fashion. Process networks reconfigure by replacing a node with a subgraph. This is only possible if the subgraph can be appropriately spliced into the incoming and outgoing edges of the original node. Here, we examine aspects of safe implementation of such computational reconfiguration (mutation of a representation), as a basis for further work on adaptive reconfiguration. [Kah74] assumes normal order (demand driven) evaluation, hence: 1. reconfiguration happens only when it is needed, and 2. no unnecessary reconfigurations occur. If an implementation is also demand driven, then (1) makes it possible to make strong statements about the process network node that is undergoing reconfiguration, and its input and output channels. Implementing reconfiguration requires that we change the process network representation dynamically; we must be able to guarantee that the mechanics of mutating the process network’s implementation structure gives the correctly formed process network before and after mutation, and that mutation must ensure that all channel connections remain correct, that no values are lost from channels, that no values are introduced or duplicated, and that the process network correctly resumes operation. We refer to this as ensuring consistency of representation during reconfiguration. Again using (1) above, and a demand driven implementation, we can state (in Fig. 1) that the (single) output channel of B is empty and hungry, and that a value must be produced on that channel immediately after the reconfiguration in order to satisfy that demand; in [Wen82], we showed how to perform the mutation safely under these conditions in both quasi-parallel and distributed process network implementation schemes. But we do not want to restrict operation of a process network to just demand driven operation; it restricts concurrency (not entirely, as a sink node can propagate demands on any number of inputs, thereby providing several concurrent activities to satisfy them, but it precludes pipelining or streaming parallelism). For this reason, we want parts of a network to proceed in a data driven fashion. This means we need to take more care in handling mutations. A C A C (a) d1 d2 (b) d1 d2 B D Figure 1: Expansion of a network from (a) to (b), replacing the existing node B with a subgraph D. B may or may not survive as part of D in the new configuration.