Global Reproducibility Through Local Control for Distributed Active Objects

Abstract

Non-determinism in a concurrent or distributed setting may lead to many different runs or executions of a program. This paper presents a method to reproduce a specific run for non-deterministic actor or active object systems. The method is based on recording traces of events reflecting local transitions at so-called stable states during execution; i.e., states in which local execution depends on interaction with the environment. The paper formalizes trace recording and replay for a basic active object language, to show that such local traces suffice to obtain global reproducibility of runs; during replay different objects may operate fairly independently of each other and in parallel, yet a program under replay has guaranteed deterministic outcome. We then show that the method extends to the other forms of non-determinism as found in richer active object languages. Following the proposed method, we have implemented a tool to record and replay runs, and to visualize the communication and scheduling decisions of a recorded run, for Real-Time ABS, a formally defined, rich active object language for modeling timed, resource-aware behavior in distributed systems.

A Survey on Parallelism and Determinism

“…Additionally, determinism properties can improve both execution support and debugging. For example, partial determinism properties of active objects have been successfully used to design recovery protocols in the context of fault-tolerance [BCDH07b], and to design deterministic replay techniques in the context of debugging support [TJS20a].…”

Active Objects with Deterministic Behaviour

“…In addition, Encore supports an advanced capability-based type system [20] which enables race-free data sharing between active objects. Confluence properties for cooperative scheduling in ABS have previously been studied, based on controlling the local scheduler [21,22].…”

From SOS to Asynchronously Communicating Actors

“…This approach is based on the notion of stable points in the execution of ABS programs [5], at which an object requires external input to make progress (either an event or a scheduling decision). The semantics of ABS then allows us to prove that executions are globally confluent at the granularity of stable points [5,6]. Consequently, it is sufficient to reason about one object at a time between stable points in the program execution.…”

“…This paper is an extended abstract of an invited talk given at FOCLASA 2019. Further details of the ideas discussed in this paper may be found in [5,6] and the source of the original SOS model which triggered our interest in this line of investigation may be found in [7].…”