From a Last-Week Add-On to First-Day Practice: Three Years of Integrating MPI Performance Analysis into HPC Education with EduMPI Suite

From a Last-Week Add-On to First-Day Practice: Three Years of Integrating MPI Performance Analysis into HPC Education with EduMPI Suite Anna-Lena Roth

∗ , Jonas Posner

∗

∗ Fulda University of Applied Sciences, Germany

arXiv:2610.06352v1 [cs.DC] 5 Oct 2026

{anna-lena.roth, jonas.posner}@cs.hs-fulda.de and open tools including CUBE [7] and TAU [8]. These tools are powerful, but effective use requires cluster operation, instrumentation, measurement, and tool-specific interpretation: students must know which performance metrics to inspect, where performance problems arise in code, and which view can confirm a hypothesis, while also handling SSH, Linux commands, file transfer, compilation, job scripts, and workload managers such as SLURM. The usual post-mortem workflow adds further friction because data become visible only after instrumentation, a measurement run, and result-file generation. In education, this combined complexity often pushes performance analysis toward the course end, where it can appear as a specialized topic rather than an integral part of MPI development. Our master’s-level Parallel Programming course at Fulda University of Applied Sciences, Germany, followed that conventional response until 2023: teach MPI first and introduce performance analysis in the final exercise. EduMPI Suite was developed to invert this sequence. It automates cluster use: students start MPI programs from EduMPI I. I NTRODUCTION GUI on their local machines and execute them on the cluster The Message Passing Interface (MPI) is the dominant without first mastering SSH, job scripts, or workload-manager programming model for distributed-memory High Performance commands. Performance analysis is triggered automatically, Computing (HPC) systems: developers decompose work across and performance data are shown in near-real-time in the processes and exchange data explicitly through point-to-point, GUI. Students are therefore confronted with concrete MPI one-sided, and collective communication. Efficient MPI pro- communication behavior while they are executing the program gramming therefore requires more than syntactically correct on a cluster, not only after a separate post-mortem workflow. code. The same program can behave differently depending EduMPI Suite presents this behavior through complementary on process placement, node topology, memory hierarchy, 2D and 3D views that connect ranks, messages, time intervals, network characteristics, and workload manager decisions. This and physical placement. As a novel feature, it also opens performance-portability problem is part of everyday parallel the black box of collective communication by visualizing the programming: experienced developers analyze applications on internal point-to-point exchanges selected by the MPI library, the cluster where they will run and use evidence from that which remain largely hidden in professional tools [9]. The system to locate waiting states, load imbalance, and com- goal is not to replace professional performance analysis, but munication volume. These competencies therefore need to be to make it available from the first MPI program onward. acquired during programming education. Current HPC curricula This paper reports three years of integrating performance accordingly recommend practical performance analysis and analysis into our Parallel Programming course and developing visualization as part of parallel-programming instruction [1], EduMPI Suite through classroom use and usability studies. We [2], [3], [4]. ask: How did moving from last-week professional-tool exercises Professional analysis workflows, however, are difficult to to first-day EduMPI-supported practice affect students’ percepintroduce when beginners first write MPI code. Score-P [5] tion and use of MPI performance analysis? We contribute (1) is an open, widely used instrumentation and measurement previously unpublished survey evidence from two historical endinfrastructure. The resulting traces and profiles can be explored of-course TAU cohorts; (2) a synthesis of the design decisions with established tools such as the commercial Vampir [6] that shaped EduMPI Suite across the three-year development Abstract—Performance analysis is essential for scalable MPI programs, yet tool and workflow complexity often relegates it to the end of Parallel Programming courses. EduMPI Suite lowers this entry barrier by automating cluster execution, measurement, and near-real-time analysis while visualizing MPI communication from the first program. This paper presents a three-year experience report on its integration into HPC education. We trace the transition of performance analysis from a last-week add-on to a recurring element of MPI development, using unpublished historical survey evidence and individual 2025 questionnaire data. Earlier cohorts perceived performance analysis as unnecessary, difficult, or unintuitive. In the final study, all students recommended introducing EduMPI Suite early in the course. Of 33 students, 29–33 rated it higher than conventional tools for understanding program behavior, interpreting communication, identifying bottlenecks, obtaining visual feedback, and navigating the tool. These findings support embedding performance analysis throughout MPI education before transitioning learners to professional environments. Index Terms—HPC Education, MPI, Performance Analysis, Visualization, Experience Report

Author accepted manuscript. Accepted for publication in the SC26 Workshops. © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.

trajectory; and (3) a previously unpublished individual itemlevel comparison from a 2025 usability study in which students evaluated EduMPI Suite against TAU or CUBE. Together, they show how EduMPI Suite strengthens performance-analysis understanding while bridging to professional environments.

instrumentation and measurement from a median of eight to one [12]. This motivated the focus on automated cluster execution and immediate near-real-time feedback. Classroom tests in WiSe 2024/25 further examined collective communication. Among 30 students, 29 recognized MPI_Bcast as more efficient than an iterative point-to-point II. E DU MPI S UITE BACKGROUND broadcast, 28 identified rank 0 as the bottleneck of the iterative A. Architecture and workflow point-to-point broadcast, and students identified message-sizeEduMPI Suite consists of three core components [10], [11]. dependent algorithm changes within MPI_Bcast [9]. A comprehensive comparative usability study in summer EduMPI is a fork of Open MPI that preserves standard MPI semester (SuSe) 2025 then evaluated the mature version of functionality while adding measurement of communication EduMPI Suite against TAU or CUBE with 33 master’s students. events. EduStore is a time-series database that receives these Students solved 96% of analysis tasks with EduMPI Suite events from running applications and supports low-latency compared with 60% using the professional tools, required queries. EduMPI GUI is the cross-platform desktop interface fewer interventions, and all 33 recommended EduMPI Suite for through which students execute programs and inspect the educational use [13]. The counterbalanced design also indicated collected data. It is available for common student operating transfer: groups using EduMPI Suite first later performed better systems, including Windows, Linux, and macOS. with TAU/CUBE than groups starting with the professional The resulting workflow hides the operational steps that tools. These studies establish the tool’s usability trajectory. normally precede performance analysis. Students select source This paper adds the longitudinal classroom interpretation, files and process counts in the GUI. EduMPI Suite transfers previously unpublished historical baseline evidence, and the the files, generates the SLURM job, compiles and launches previously unpublished individual item-level comparison from the application, and monitors its state on the cluster. While the SuSe 2025 study. the program is running, EduMPI streams MPI event data to EduStore and EduMPI GUI queries them for near-real-time III. F ROM A DD -O N TO F IRST-DAY P RACTICE visualization. For responsiveness in classroom use, EduMPI A. The baseline workflow GUI primarily shows data aggregated per process and second. In SuSe 2022 and WiSe 2022/23, performance analysis Students can select intervals, abort runs, review experiments, appeared at the end of the course, after students had already and automatically generate Score-P profiles for professional completed most MPI programming tasks [14]. Students invisualization tools [10], [12]. Evaluated runtime overhead stalled TAU locally, instrumented an MPI application with remains below 3.8%, and the architecture has supported Score-P, submitted cluster jobs, transferred result files, and concurrent classroom use [11]. analyzed them with TAU after an in-class demonstration. This B. MPI communication as the central representation baseline required novices to solve operational, representational, EduMPI Suite was designed around MPI communication and conceptual problems simultaneously. Visualization was rather than around general-purpose profiling tables. Its central available only after a successful post-mortem pipeline, while representation connects source-level MPI behavior to physical learners still had to know which metric or view to inspect. In SuSe 2022, only 6 of 17 respondents rated TAU’s cluster structure. A 3D overview places processes inside nodes and visualizes communication partners by connecting them navigation intuitive, 7 agreed that TAU helped them understand with arrows, while 2D views expose detailed data volumes. program behavior, and 6 independently identified improvement Runtime options represent send/receive ratios across processes potential (Fig. 1). Just 3 agreed that they understood codeinstrumentation principles, while 5 considered local execution as well as waiting states. The most distinctive feature is the visualization of collective sufficient and cluster use unnecessary. Perceived relevance did communication. While professional tools display MPI_Bcast not reliably translate into usable analysis or cluster practice. Only 3 of 14 rated TAU’s navigation intuitive, 4 indepenas a collective event, EduMPI Suite exposes MPI’s underlying point-to-point messages [9]. This makes hidden implementation dently identified improvement potential of an MPI program, decisions observable: students can compare a manually imple- and 2 agreed that they understood the instrumentation and mented broadcast with MPI_Bcast, identify bottlenecks, and measurement workflow. At the same time, 10 wanted earlier explain differences in communication structure and scalability. analysis and 8 intended to reuse TAU. Across the three shared dimensions, positive ratings remained a minority in both C. Prior evaluations cohorts: navigation (6/17 and 3/14), instrumentation (3/17 and EduMPI Suite was refined through several user and usability 2/14), and independent identification of improvement potential tests. In winter semester (WiSe) 2023/24, a preliminary usability (6/17 and 4/14). Late placement generated interest but left study with a conventional Score-P/Scalasca/CUBE workflow insufficient time to build independent competence. In the WiSe showed that EduMPI Suite reduced the median time to the 2023/24 conventional-tool study, six students described analysis first visualization from about 26 minutes to about 5 minutes as unnecessary or too difficult and suggested it, at most, as a and reduced moderator interventions during performance bonus or final topic [12].

1

2

3

4

Summer semester 2022 (n=17) 6

Intuitive navigation

Understand behavior

Cluster use unnecessary

3

7

4

2

7

3

5

7

Should be introduced earlier

Understand instrumentation

Identify improvement potential

4

5

Intuitive navigation

5

6

Understand instrumentation

Identify improvement potential

6

7

1 = low/disagree; 5 = high/agree

Winter semester 2022/23 (n=14)

5

3

5

Would use TAU in future

5

2

3

5

3

5

6

4

3

5

5

2

3

6

2

Figure 1. Likert distributions from two end-of-course TAU surveys. Numbers inside segments are response counts; unlabeled segments contain one response.

B. Evolution of the intervention

evaluation-to-design loop. The revised course uses EduMPI Suite from the first MPI program onward so that students learn to identify bottlenecks, waiting states, and late senders and relate them to their code and MPI communication. It lowers the entry barrier by deferring SSH, SLURM, manual execution, and professional measurement workflows until later in the course, when students transition to CUBE or TAU for post-mortem analysis.

EduMPI Suite was developed in response to the limitations of the conventional sequence used in SuSe 2022 and WiSe 2022/23, in which students first learned MPI programming and encountered TAU only in the final exercise. These limitations motivated a first major design decision: performance analysis should become an integral part of learning MPI rather than remain a final add-on. Accordingly, EduMPI Suite was designed IV. R ESULTS A FTER T HREE Y EARS to reduce operational barriers such as SSH access, job scripts, and workload managers, while providing immediate, near-real- A. From group-level findings to classroom experience time feedback from the first MPI program onward. After three years, the most important outcome is that The next iterations addressed barriers exposed by class- performance analysis moved from a final exercise to a repeated room use and usability studies. The WiSe 2023/24 Score- way of reasoning about MPI programs. In the published P/Scalasca/CUBE study showed that conventional workflows comprehensive usability study, students solved 96% of analysis consumed substantial time and instructor support before a tasks with EduMPI Suite compared with 60% using TAU first visualization [12]. EduMPI Suite therefore automated or CUBE; interventions were needed in 29% versus 55% cluster access, job generation, measurement, and near-real- of tasks; and EduMPI Suite was usually faster for locating time visualization. Once this made early use possible, the processes, communication patterns, or bottlenecks [13]. All 33 problem shifted from access to explanation: students needed participants recommended EduMPI Suite for educational use, views that connected source-level MPI calls to ranks, nodes, compared with one unqualified recommendation for TAU or cores, communication volume, and phases of execution. This CUBE. The crossover design further showed a transfer effect: motivated the timeline, time selection, topology-aware 2D/3D groups that used EduMPI Suite first solved more subsequent views, communication arrows, and per-rank data later described TAU/CUBE tasks correctly and needed fewer interventions as core EduMPI Suite features [10]. with those professional tools. The user tests focusing on collective communication exEduMPI Suite was strongest when students connected a tended this idea from visible user-level messages to MPI- visible symptom to an MPI mechanism, such as waiting ranks, internal behavior. Because professional tools usually present centralized communication, process placement, or collective collectives as library calls, students cannot easily see why algorithms. Its main weakness was showing summarized MPI_Bcast differs from a manually implemented broadcast statistics about the whole program: because EduMPI Suite or why message size changes the algorithm selected. EduMPI is optimized for near-real-time, time-segmented feedback, Suite was therefore extended to expose the internal point- CUBE and TAU remain stronger for post-mortem profiles. We to-point exchanges used by the MPI implementation behind therefore treat EduMPI Suite as the first scaffold for forming collectives, making its selected communication pattern visible; hypotheses, followed by professional tools for aggregated postthe teaching tasks were redesigned around explaining these mortem analysis. algorithmic patterns [9]. Finally, the comprehensive 2025 usability study showed both B. Previously unpublished individual comparison the strength and the boundary of this scaffold: EduMPI Suite The published comprehensive usability study reported groupsupported local diagnosis and mental-model formation, while level task success, interventions, processing time, recommenCUBE and TAU remained valuable for aggregated post-mortem dations, and transfer observations [13]. It did not present the performance analysis [13]. Table I summarizes the resulting individual post-test questionnaire data [15]. For this experience

Table I E VOLUTION OF THE CLASSROOM INTERVENTION AND EVIDENCE BASE . Phase

Classroom intervention

Observed barrier or opportunity

Resulting design/teaching decision

SuSe 2022 / TAU in the final exercise WiSe 2022/23

Move from one final TAU session to recurring performance-analysis tasks throughout the course; remove operational barriers

WiSe 2023/24

Automate cluster login, job generation, measurement, and visualization in the GUI; support repeated runs [12] Add timeline/time selection and topology-aware 2D/3D views with communication arrows and per-rank data [10]

SuSe 2024

WiSe 2024/25

SuSe 2025

Guided TAU use was possible and partly valued, but analysis remained a late specialist topic and students rarely identified improvement potential without support Preliminary usability study: Conventional workflow required much more time Score-P/Scalasca/CUBE vs. EduMPI Suite and command-level help before first visualization EduMPI Suite during regular MPI Early execution worked, but explanations needed exercises clearer links among code, ranks, nodes/cores, and phases User studies on collective communication Students needed evidence for why collectives outperform hand-written loops and how MPI changes algorithms Comprehensive, counterbalanced EduMPI EduMPI Suite supported local diagnosis; Suite–TAU/CUBE usability study professional tools remained useful for aggregation and detailed post-run views

Expose internal collective point-to-point patterns and design tasks around explaining algorithm changes [9] Position EduMPI Suite as scaffold: start with visual hypotheses, then export/transfer to CUBE or TAU and add summaries [13]

report, we use these individual ratings to show whether the The second change concerns support needs and transfer. favorable experience was concentrated in a few groups or Beyond students’ preference for EduMPI Suite, the higher visible across students. The 33 participants worked in dyads or task success, lower support needs, and subsequent TAU/CUBE triads in a counterbalanced design: six groups used EduMPI results reported above support the teaching pattern that emerged Suite before TAU or CUBE, and six groups used the reverse from the three-year trajectory: first make evidence immediate, order; nine groups selected CUBE and three selected TAU. then build a visual communication model, reveal collective This resulted in 25 individual comparisons between EduMPI algorithms, and finally move to professional tools once students Suite and CUBE and eight between EduMPI Suite and TAU. can formulate performance questions. The distributions in Fig. 2 show that the result is not driven The instructor’s role changes accordingly. Because students by one professional tool or by a small number of enthusiastic need less procedural help with cluster access, measurement, and groups. In the CUBE subgroup, all 25 participants rated tool operation, the instructor can focus on conceptual questions: EduMPI Suite 4 or 5 for understanding program behavior and Which rank waits, how does doubling ranks change the tool navigation, whereas 24 rated CUBE between 1 and 3 on pattern, and why does MPI select different broadcast algorithms each item. In the TAU subgroup, all eight rated EduMPI Suite for different message sizes? Screenshots and time selections 4 or 5 on every dimension; TAU received only one rating above become evidence in explanations, making performance analysis 3. The paired direction is equally consistent: every participant assessable as reasoning rather than tool operation. rated EduMPI Suite higher for understanding program behavior, In practice, this changed the course culture. Performance interpreting communication, and tool navigation. For bottleneck analysis became a regular part of MPI programming: students identification, 29 rated EduMPI Suite higher and four tied. For increasingly asked not only whether a program was correct, but visual feedback, only one participant rated the professional tool how well it behaved and where performance could be improved. higher. These individual data strengthen the experience-report Several students used screenshots from performance-analysis claim: the shift was not merely that groups completed more tools, especially EduMPI Suite, in final presentations to explain tasks, but that students individually perceived EduMPI Suite their programs and demonstrate observed communication as the more helpful representation for the core activities of behavior. The near-real-time visualization supported more beginner-level MPI performance analysis. independent cluster use: when students discovered a problem early, they could abort a run and free limited educational cluster resources instead of waiting for post-mortem results. V. W HAT C HANGED C OMPARED WITH THE BASELINE The approach also has limitations. While EduMPI Suite helps In the late TAU baseline, positive ratings remained a minority students relate visible symptoms to MPI mechanisms, its focus for intuitive navigation (6/17 in 2022 and 3/14 in 2022/23), un- on near-real-time, time-segmented feedback makes aggregated derstanding instrumentation (3/17 and 2/14), and independently profile overviews less accessible. Questions such as which MPI identifying improvement potential (6/17 and 4/14). In 2025, all call dominates the total runtime still motivate a handover to 33 participants rated EduMPI Suite positively for understanding CUBE or TAU. Automation also did not make every represenprogram behavior and tool navigation, and every participant tation intuitive: some students still criticized 3D navigation, rated it higher than TAU/CUBE for understanding program ambiguous colors, and missing labels. Finally, EduMPI Suite rebehavior, interpreting communication, and tool navigation. The quires deployment effort: EduStore, cluster integration, and oncomparison is not a controlled cohort study because wording, going maintenance. The architecture has been reliable with up tasks, and timing changed, but the direction is consistent: to 24 simultaneous classroom users and low overhead [11], but the activity that had previously remained late and support- the evidence remains a single-institution experience with small dependent became understandable enough to be used repeatedly historical cohorts and a 2025 study shaped by group work, prior during MPI development. EduMPI Suite exposure, and the selected comparison tools.

1

2

3

4

5

EduMPI Suite vs. Cube (paired subgroup, n=25) Understand behavior

EduMPI Cube

13 6

Interpret communication

EduMPI Cube

10

Identify bottlenecks

EduMPI

6

Cube

12 11

Visual feedback helpful

EduMPI Cube

7

Intuitive navigation

EduMPI

8

Cube

7

Interpret communication

EduMPI

Identify bottlenecks

EduMPI

17 10

Visual feedback helpful

EduMPI

Intuitive navigation

EduMPI

TAU

6 17

12

3 7

TAU

5

1

5

3

2

6 4

TAU

10

7

5

TAU

4

17 7

EduMPI TAU

7

11

7

Understand behavior

12 11

1 = low/disagree; 5 = high/agree

EduMPI Suite vs. TAU (paired subgroup, n=8)

4

3

2

2

1

6 1

2 5

2

5 3

3 3

2

Figure 2. Previously unpublished individual paired Likert distributions from the SuSe 2025 evaluation, separated by the professional tool selected by each group. Each participant rated EduMPI Suite and CUBE or EduMPI Suite and TAU on the same five items. Numbers inside segments are response counts; unlabeled narrow segments contain one or two responses.

For adoption elsewhere, the main lesson is to reproduce this sequence rather than treat EduMPI Suite as a replacement for professional tools. Institutions must plan for this infrastructure and its maintenance, while retaining CUBE or TAU for aggregate post-mortem questions. Less cluster-experienced populations may benefit especially from separating operational setup from performance reasoning; whether the same scaffold benefits other institutions or student populations, however, requires evaluation beyond our master’s-level setting. VI. C ONCLUSION Three years of classroom use changed performance analysis from a last-week add-on into first-day practice. In the original setting, operational difficulty weakened the perceived relevance of the topic: some students considered it unnecessary, and only a minority independently found optimization potential with TAU. With EduMPI Suite embedded throughout MPI practice, performance evidence became expected feedback. Students used it to reason about program behavior, motivate improvements, and present their results, and all 33 participants in the latest evaluation recommended the tool for education. The two historical cohorts document the original tension between interest, perceived importance, and independent use. Students valued guided work with TAU and wanted analysis earlier, yet few independently identified optimization potential and several reported substantial barriers. The 2025 item-level comparison shows a uniform advantage for the representational functions needed in the scaffold. The feedback-to-transfer pattern explains how this change can be reproduced without making EduMPI Suite the final destination: immediate evidence, a visual communication model, explicit collective algorithms, and professional tools once learners can formulate questions. Future studies should quantify the effect more precisely and test delayed transfer, but the present longitudinal evidence supports moving performance analysis earlier and making it recurring MPI practice.

ACKNOWLEDGMENT The authors used AI assistants to help polish the wording, grammar, and clarity of the manuscript. All technical content, including the research ideas, claims, methodology, data analysis, and results, was produced and verified by the authors. R EFERENCES [1] [2]

[3]

[4]

[5]

[6] [7]

[8] [9]

[10]

[11]

[12]

[13]

[14] [15]

P. Bouvry et al., “The European master for HPC curriculum,” J. Parallel Distributed Comput., 2025. DOI: 10.1016/J.JPDC.2025.105081 S. K. Prasad et al., NSF/IEEE-TCPP Curriculum Initiative on Parallel and Distributed Computing: Core Topics for Undergraduates (Version 2.0-beta), 2020. [Online]. Available: https://tcpp.cs.gsu.edu/curriculum/?q=home R. K. Raj et al., “High Performance Computing Education: Current Challenges and Future Directions,” in Proceedings of the Working Group Reports on Innovation and Technology in Computer Science Education, ACM, 2020. DOI: 10.1145/ 3437800.3439203 A.-L. Roth and T. Süß, “Performance Analysis Tools for MPI Applications and their Use in Programming Education,” in Companion of the 2023 ACM/SPEC International Conference on Performance Engineering, ser. ICPE ’23 Companion, ACM, 2023. DOI: 10.1145/3578245.3584358 A. Knüpfer et al., “Score-P: A Joint Performance Measurement Run-Time Infrastructure for Periscope, Scalasca, TAU, and Vampir,” in Tools for High Performance Computing, Springer, 2012. DOI: 10.1007/978-3-642-31476-6_7 A. Knüpfer et al., “The Vampir Performance Analysis Tool-Set,” in Tools for High Performance Computing, Springer, 2008. DOI: 10.1007/978-3-540-68564-7_9 P. Saviankou, M. Knobloch, A. Visser, and B. Mohr, “Cube v4: From Performance Report Explorer to Performance Analysis Tool,” Procedia Computer Science, 2015. DOI : 10.1016/j.procs.2015.05.320 S. Shende and A. D. Malony, “The TAU Parallel Performance System,” Int. J. High Perform. Comput. Appl., 2006. DOI: 10.1177/1094342006064482 A.-L. Roth, D. James, M. Kuhn, and D. Frisch, “Making MPI Collective Operations Visible: Understanding Their Utility and Algorithmic Insights,” in European Conference on Parallel and Distributed Processing (Euro-Par), Dresden, Germany, Springer, 2025. DOI: 10.1007/978-3-031-99854-6\_5 A.-L. Roth, D. James, and M. Kuhn, “EduMPI - Simplifying the Use of HighPerformance Clusters and Focusing Performance Analysis in Parallel Programming Education,” PARS-Mitteilungen, 2025. [Online]. Available: https://dl.gi.de/handle/ 20.500.12116/46131 A.-L. Roth, J. Posner, and M. Kuhn, “Enabling Interactive Visualization for Near-Real-Time Performance Analysis of MPI Applications,” in International Conference on Performance Engineering, ACM, 2026. DOI: 10.1145/3777911. 3800642 A.-L. Roth, D. James, M. Kuhn, and J. Konert, “Enhancing Parallel Programming Education with High-Performance Clusters Utilizing Performance Analysis,” in Fachtagung Bildungstechnologien der Gesellschaft für Informatik e.V., DELFI, Fulda, Germany, vol. P-356, GI, 2024. DOI: 10.18420/delfi2024\_42 A.-L. Roth, D. James, J. Posner, and M. Kuhn, “Performance Analysis in Parallel Programming Education: A Comparative Usability Study,” PARS-Mitteilungen, 2026. [Online]. Available: https://dl.gi.de/handle/20.500.12116/48841 A.-L. Roth, End-of-Course Student Surveys on MPI Performance Analysis with TAU in a Master’s Parallel Programming Course, 2026. DOI: 10.5281/zenodo.22662192 A.-L. Roth, Usability Evaluation EduMPI 2025, 2025. DOI: 10.5281/zenodo. 16798981

Related documents

• From a Last-Week Add-On to First-Day Practice: Three Years of Integrating MPI Performance Analysis into HPC Education with EduMPI Suite#1202771
• Compiler-Assisted Performance Analysis Workflows for HPC Applications / Sebastian Kreutzer ; Gutachter: Christian Bischof, Jesus Labarta Mancho#1145408
• HPCAgent-Bench: A Framework for Sound Benchmarking of Code-Optimizing Agents, Harnesses, and Skills#1355662
• HPCAgent-Bench: A Framework for Sound Benchmarking of Code-Optimizing Agents, Harnesses, and Skills#1348926
• Compiler-based code generation and autotuning for geometric multigrid on GPU-accelerated supercomputers#614777