Deterministic Replay

Abstract

Deterministic replay is a type of emerging technique dedicated to providing deterministic executions of computer programs in the presence of nondeterministic factors. The application scopes of deterministic replay are very broad, making it an important research topic in domains such as computer architecture, operating systems, parallel computing, distributed computing, programming languages, verification, and hardware testing. In this survey, we comprehensively review existing studies on deterministic replay by introducing a taxonomy. Basically, existing deterministic replay schemes can be classified into two categories, single-processor (SP) schemes and multiprocessor (MP) schemes. By reviewing the details of these two categories of schemes respectively, we summarize and compare how existing schemes address technical issues such as log size, record slowdown, replay slowdown, implementation cost, and probe effect, which may shed some light on future studies on deterministic replay.

Efficient and deterministic record & replay for actor languages

“…This makes it unsuitable for non-deterministic programs, where the occurrence of a bug may depend on a rare scheduling of messages. As mentioned before, record & replay [8] enables deterministic re-execution of a previously recorded program execution, and thereby enables cyclic debugging also for non-deterministic programs. During the initial execution, such approaches record a program trace, which is then used during replay to guide the execution and reproduce, for instance, input from external sources and scheduling decisions, and thereby eliminate all non-determinism.…”

“…Record & replay for parallel and concurrent programs has been studied before, but a majority of the previous work focused on shared memory concurrency and MPI-like message passing [8]. Recent work focused either on single event loops or did not consider performance [2,6,35].…”

Abstractions for Practical Virtual Machine Replay

“…Several surveys provide taxonomies of early [20] and more recent work [16,30] in the area of replay debugging. Chen et al provide a good overview of recent approaches to multiprocessor replay [13]. Contemporary notions of system replay originated as parts of distributed checkpoint protocols [6], replay debuggers for parallel systems [18], and fault-tolerant replication approaches [8].…”

“…Numerous applications of deterministic replay were suggested: e.g., debugging and analysis of complex software systems [15,26,27,32,33,35,40,41], fault-tolerant replication [9,43,44], performance analysis [4], and forensics [11,19,22]). A number of deterministic replay systems were developed along with advanced techniques for reconstructing execution of parallel [1,12,13,38] and distributed systems. However, despite academic success, deterministic replay did not become a de facto part of systems and virtualization stacks.…”

“…These are appropriate for many production use-cases where it is often acceptable to obtain scalability by running multiple guests on a powerful machine. A number of promising research efforts have addressed the problem of high overhead [13,23] in recording multiprocessor guests. Still, these techniques require assumptions that are often unacceptable for production environments: e.g., the need to tolerate the overheads of whole-system binary translation [12], heavyweight execution-reconstruction techniques that are limited to several seconds of recording [1,38], the inability to record a whole system due to strict limits on the amount of sharedmemory nondeterminism [33], intrusive changes to the entire OS stack [7], or specialized hardware [29].…”