Drechsler, Joscha (2019)
Concurrency and Distribution in Reactive Programming.
Technische Universität Darmstadt
Dissertation, Erstveröffentlichung
Kurzbeschreibung (Abstract)
Distributed Reactive Programming is a paradigm for implementing distributed interactive applications modularly and declaratively. Applications are defined as dynamic distributed dataflow graphs of reactive computations that depend upon each other, similar to formulae in spreadsheets. A runtime library propagates input changes through the dataflow graph, recomputing the results of affected dependent computations while adapting the dataflow graph topology to changing dependency relations on the fly. Reactive Programming has been shown to improve code quality, program comprehension and maintainability over modular interactive application designs based on callbacks. Part of what makes Reactive Programming easy to use is its synchronous change propagation semantics: Changes are propagated such that no computation can ever observe an only partially updated state of the dataflow graph, i.e., each input change together with all dependent recomputations appears to be instantaneous.
Traditionally, in local single-threaded applications, synchronous semantics are achieved through glitch freedom consistency: a recomputation may be executed only after all its dependencies have been recomputed. In distributed applications though, this established approach breaks in two ways. First, glitch freedom implies synchronous semantics only if change propagations execute in isolation. Distributed applications are inherently exposed to concurrency in that multiple threads may propagate different changes concurrently. Ensuring isolation between concurrent change propagations requires the integration of additional algorithms for concurrency control. Second, applications’ dataflow graphs are spread across multiple hosts. Therefore, distributed reactive programming requires algorithms for both glitch freedom and concurrency control that are decentralized, i.e., function without access to shared memory. A comprehensive survey of related prior works shows, that none have managed to solve this issue so far.
This dissertation introduces FrameSweep, the first algorithm to provide synchronous propagation semantics for concurrent change propagation over dynamic distributed dataflow graphs. FrameSweep provides isolation and glitch freedom through a combination of fine-grained concurrency control algorithms for linearizability and serializability from databases with several aspects of change propagation algorithms from Reactive Programming and Automatic Incrementalization. The correctness of FrameSweep is formally proven through the application of multiversion concurrency control theory. FrameSweep decentralizes all its algorithmic components, and can therefore be integrated into any reactive programming library in a syntactically transparent manner: Applications can continue to use the same syntax for Reactive Programming as before, without requiring any adaptations of their code. FrameSweep therefore provides the exact same semantics as traditional local single-threaded Reactive Programming through the exact same syntax. As a result, FrameSweep proves by example that interactive applications can reap all benefits of Reactive Programming even if they are concurrent or distributed.
A comprehensive empirical evaluation measures through benchmarks, how FrameSweep’s performance and scalability are affected by a multitude of factors, e.g., thread contention, dataflow topology, dataflow topology changes, or distribution topology. It shows that FrameSweep’s performance compares favorably to alternative scheduling approaches with weaker guarantees in local applications. An existing Reactive Programming application is migrated to FrameSweep to empirically verify the claims of semantic and syntactic transparency.
Typ des Eintrags: | Dissertation | ||||
---|---|---|---|---|---|
Erschienen: | 2019 | ||||
Autor(en): | Drechsler, Joscha | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | Concurrency and Distribution in Reactive Programming | ||||
Sprache: | Englisch | ||||
Referenten: | Mezini, Prof. Dr. Mira ; De Meuter, Prof. Dr. Wolfgang | ||||
Publikationsjahr: | 2019 | ||||
Ort: | Darmstadt | ||||
Datum der mündlichen Prüfung: | 11 Februar 2019 | ||||
URL / URN: | https://tuprints.ulb.tu-darmstadt.de/5228 | ||||
Kurzbeschreibung (Abstract): | Distributed Reactive Programming is a paradigm for implementing distributed interactive applications modularly and declaratively. Applications are defined as dynamic distributed dataflow graphs of reactive computations that depend upon each other, similar to formulae in spreadsheets. A runtime library propagates input changes through the dataflow graph, recomputing the results of affected dependent computations while adapting the dataflow graph topology to changing dependency relations on the fly. Reactive Programming has been shown to improve code quality, program comprehension and maintainability over modular interactive application designs based on callbacks. Part of what makes Reactive Programming easy to use is its synchronous change propagation semantics: Changes are propagated such that no computation can ever observe an only partially updated state of the dataflow graph, i.e., each input change together with all dependent recomputations appears to be instantaneous. Traditionally, in local single-threaded applications, synchronous semantics are achieved through glitch freedom consistency: a recomputation may be executed only after all its dependencies have been recomputed. In distributed applications though, this established approach breaks in two ways. First, glitch freedom implies synchronous semantics only if change propagations execute in isolation. Distributed applications are inherently exposed to concurrency in that multiple threads may propagate different changes concurrently. Ensuring isolation between concurrent change propagations requires the integration of additional algorithms for concurrency control. Second, applications’ dataflow graphs are spread across multiple hosts. Therefore, distributed reactive programming requires algorithms for both glitch freedom and concurrency control that are decentralized, i.e., function without access to shared memory. A comprehensive survey of related prior works shows, that none have managed to solve this issue so far. This dissertation introduces FrameSweep, the first algorithm to provide synchronous propagation semantics for concurrent change propagation over dynamic distributed dataflow graphs. FrameSweep provides isolation and glitch freedom through a combination of fine-grained concurrency control algorithms for linearizability and serializability from databases with several aspects of change propagation algorithms from Reactive Programming and Automatic Incrementalization. The correctness of FrameSweep is formally proven through the application of multiversion concurrency control theory. FrameSweep decentralizes all its algorithmic components, and can therefore be integrated into any reactive programming library in a syntactically transparent manner: Applications can continue to use the same syntax for Reactive Programming as before, without requiring any adaptations of their code. FrameSweep therefore provides the exact same semantics as traditional local single-threaded Reactive Programming through the exact same syntax. As a result, FrameSweep proves by example that interactive applications can reap all benefits of Reactive Programming even if they are concurrent or distributed. A comprehensive empirical evaluation measures through benchmarks, how FrameSweep’s performance and scalability are affected by a multitude of factors, e.g., thread contention, dataflow topology, dataflow topology changes, or distribution topology. It shows that FrameSweep’s performance compares favorably to alternative scheduling approaches with weaker guarantees in local applications. An existing Reactive Programming application is migrated to FrameSweep to empirically verify the claims of semantic and syntactic transparency. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
URN: | urn:nbn:de:tuda-tuprints-52284 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik | ||||
Fachbereich(e)/-gebiet(e): | 20 Fachbereich Informatik 20 Fachbereich Informatik > Softwaretechnik |
||||
Hinterlegungsdatum: | 11 Aug 2019 19:55 | ||||
Letzte Änderung: | 11 Aug 2019 19:55 | ||||
PPN: | |||||
Referenten: | Mezini, Prof. Dr. Mira ; De Meuter, Prof. Dr. Wolfgang | ||||
Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 11 Februar 2019 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Frage zum Eintrag |
Optionen (nur für Redakteure)
Redaktionelle Details anzeigen |