Towards Automated Recommendation of Steady-State Detection Algorithms for Software Performance Benchmarks

Steady-state detection is a central step in software performance benchmarking, yet different detection methods rely on different assumptions and may perform best on different execution profiles. This paper formulates steady-state detection as an algorithm-recommendation problem: given a benchmark time series, predict which of two complementary detectors is more appropriate. The first, CP-SSD, identifies steady state through changepoint-based segmentation of the execution trace, whereas the second, KB-KSSD, uses a kernel-based approach to identify stable behaviour. We propose a data-driven recommender that characterises benchmark time series using automatically extracted features and combines supervised prediction with information derived from groups of time series exhibiting similar behaviour. The approach is evaluated on 407 labelled benchmark time series for which one of the two detectors can be identified as preferable with respect to a reference steady-state index. The proposed recommender selects the better detector correctly for 72.97% of the series, compared with 62.65% when always selecting the detector that performs best most often, corresponding to an absolute improvement of 10.32 percentage points. Paired statistical testing and bootstrap confidence intervals support this improvement. A comparison with a recommender using time-series features alone shows that the cluster-derived information mainly improves interpretability rather than predictive accuracy. Recommendations also remain highly stable when the unsupervised representation is learned from either the labelled subset or the full collection of 5,860 time series. These results demonstrate the feasibility of per-series steady-state detector recommendation while highlighting the need for broader labelled data, particularly for execution profiles favouring the changepoint-based detector.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-12
DOI
https://doi.org/10.48550/arxiv.2611.16566
Primary Topic
Software System Performance and Reliability
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Towards Automated Recommendation of Steady-State Detection Algorithms for Software Performance Benchmarks

Martin Beseda, Vittorio Cortellessa, Luca Traini, Federico Di Menna et al.
arXiv (Cornell University)
Software System Performance and Reliability
preprint

Towards Automated Recommendation of Steady-State Detection Algorithms for Software Performance Benchmarks

Martin Beseda, Vittorio Cortellessa, Luca Traini, Federico Di Menna, Daniele Di Pompeo
preprint en

Abstract

Steady-state detection is a central step in software performance benchmarking, yet different detection methods rely on different assumptions and may perform best on different execution profiles. This paper formulates steady-state detection as an algorithm-recommendation problem: given a benchmark time series, predict which of two complementary detectors is more appropriate. The first, CP-SSD, identifies steady state through changepoint-based segmentation of the execution trace, whereas the second, KB-KSSD, uses a kernel-based approach to identify stable behaviour. We propose a data-driven recommender that characterises benchmark time series using automatically extracted features and combines supervised prediction with information derived from groups of time series exhibiting similar behaviour. The approach is evaluated on 407 labelled benchmark time series for which one of the two detectors can be identified as preferable with respect to a reference steady-state index. The proposed recommender selects the better detector correctly for 72.97% of the series, compared with 62.65% when always selecting the detector that performs best most often, corresponding to an absolute improvement of 10.32 percentage points. Paired statistical testing and bootstrap confidence intervals support this improvement. A comparison with a recommender using time-series features alone shows that the cluster-derived information mainly improves interpretability rather than predictive accuracy. Recommendations also remain highly stable when the unsupervised representation is learned from either the labelled subset or the full collection of 5,860 time series. These results demonstrate the feasibility of per-series steady-state detector recommendation while highlighting the need for broader labelled data, particularly for execution profiles favouring the changepoint-based detector.

arXiv (Cornell University)
Software System Performance and Reliability
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon