Lehr, Jan-Patrick (2022)
From Valid Measurements to Valid Mini-Apps.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00020943
Dissertation, Erstveröffentlichung, Verlagsversion
Kurzbeschreibung (Abstract)
In high-performance computing, performance analysis, tuning, and exploration are relevant throughout the life cycle of an application. State-of-the-art tools provide capable measurement infrastructure, but they lack automation of repetitive tasks, such as iterative measurement-overhead reduction, or tool support for challenging and time-consuming tasks, e.g., mini-app creation. In this thesis, we address this situation with (a) a comparative study on overheads introduced by different tools, (b) the tool PIRA for automatic instrumentation refinement, and (c) a tool-supported approach for mini-app extraction. In particular, we present PIRA for automatic iterative performance measurement refinement. It performs whole-program analysis using both source-code and runtime information to heuristically determine where in the target application measurement hooks should be placed for a low-overhead assessment. At the moment, PIRA offers a runtime heuristic to identify compute-intensive parts, a performance-model heuristic to identify scalability limitations, and a load imbalance detection heuristic. In our experiments, PIRA compared to Score-P’s built-in filtering significantly reduces the runtime overhead in 13 out of 15 benchmark cases and typically introduces a slowdown of < 10 %. To provide PIRA with the required infrastructure, we develop MetaCG — an extendable lightweight whole-program call-graph library for C/C++. The library offers a compiler-agnostic call-graph (CG) representation, a Clang-based tool to construct a target’s CG, and a tool to validate the structure of the MetaCG. In addition to its use in PIRA, we show that whole-program CG analysis reduces the number of allocation to track by the memory tracking sanitizer TypeART by up to a factor of 2,350×. Finally, we combine the presented tools and develop a tool-supported approach to (a) identify, and (b) extract relevant application regions into representative mini-apps. Therefore, we present a novel Clang-based source-to-source translator and a type-safe checkpoint-restart (CPR) interface as a common interface to existing MPI-parallel CPR libraries. We evaluate the approach by extracting a mini-app of only 1,100 lines of code from an 8.5 million lines of code application. The mini-app is subsequently analyzed, and maintains the significant characteristics of the original application’s behavior. It is then used for tool-supported parallelization, which led to a speed-up of 35 %. The software presented in this thesis is available at https://github.com/tudasc.
Typ des Eintrags: | Dissertation | ||||
---|---|---|---|---|---|
Erschienen: | 2022 | ||||
Autor(en): | Lehr, Jan-Patrick | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | From Valid Measurements to Valid Mini-Apps | ||||
Sprache: | Englisch | ||||
Referenten: | Bischof, Prof. Dr. Christian ; Schulz, Prof. Dr. Martin ; Chandrasekaran, Prof. PhD. Sunita | ||||
Publikationsjahr: | 2022 | ||||
Ort: | Darmstadt | ||||
Kollation: | viii, 139 Seiten | ||||
Datum der mündlichen Prüfung: | 3 September 2021 | ||||
DOI: | 10.26083/tuprints-00020943 | ||||
URL / URN: | https://tuprints.ulb.tu-darmstadt.de/20943 | ||||
Kurzbeschreibung (Abstract): | In high-performance computing, performance analysis, tuning, and exploration are relevant throughout the life cycle of an application. State-of-the-art tools provide capable measurement infrastructure, but they lack automation of repetitive tasks, such as iterative measurement-overhead reduction, or tool support for challenging and time-consuming tasks, e.g., mini-app creation. In this thesis, we address this situation with (a) a comparative study on overheads introduced by different tools, (b) the tool PIRA for automatic instrumentation refinement, and (c) a tool-supported approach for mini-app extraction. In particular, we present PIRA for automatic iterative performance measurement refinement. It performs whole-program analysis using both source-code and runtime information to heuristically determine where in the target application measurement hooks should be placed for a low-overhead assessment. At the moment, PIRA offers a runtime heuristic to identify compute-intensive parts, a performance-model heuristic to identify scalability limitations, and a load imbalance detection heuristic. In our experiments, PIRA compared to Score-P’s built-in filtering significantly reduces the runtime overhead in 13 out of 15 benchmark cases and typically introduces a slowdown of < 10 %. To provide PIRA with the required infrastructure, we develop MetaCG — an extendable lightweight whole-program call-graph library for C/C++. The library offers a compiler-agnostic call-graph (CG) representation, a Clang-based tool to construct a target’s CG, and a tool to validate the structure of the MetaCG. In addition to its use in PIRA, we show that whole-program CG analysis reduces the number of allocation to track by the memory tracking sanitizer TypeART by up to a factor of 2,350×. Finally, we combine the presented tools and develop a tool-supported approach to (a) identify, and (b) extract relevant application regions into representative mini-apps. Therefore, we present a novel Clang-based source-to-source translator and a type-safe checkpoint-restart (CPR) interface as a common interface to existing MPI-parallel CPR libraries. We evaluate the approach by extracting a mini-app of only 1,100 lines of code from an 8.5 million lines of code application. The mini-app is subsequently analyzed, and maintains the significant characteristics of the original application’s behavior. It is then used for tool-supported parallelization, which led to a speed-up of 35 %. The software presented in this thesis is available at https://github.com/tudasc. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
Status: | Verlagsversion | ||||
URN: | urn:nbn:de:tuda-tuprints-209439 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik | ||||
Fachbereich(e)/-gebiet(e): | 20 Fachbereich Informatik 20 Fachbereich Informatik > Scientific Computing |
||||
Hinterlegungsdatum: | 30 Mär 2022 13:48 | ||||
Letzte Änderung: | 31 Mär 2022 07:55 | ||||
PPN: | |||||
Referenten: | Bischof, Prof. Dr. Christian ; Schulz, Prof. Dr. Martin ; Chandrasekaran, Prof. PhD. Sunita | ||||
Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 3 September 2021 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Frage zum Eintrag |
Optionen (nur für Redakteure)
Redaktionelle Details anzeigen |