Thread Scaling of Hexaly on the Time-Dependent VRPTW: Raw Experiment Data

Raw run records for the measurement report "Thread Scaling of Hexaly on the TDVRPTW across Two Model Encodings" (Florian Rascoussier, 2026). Version 2 of this dataset accompanies version 2 of the report: it contains the 540 solver runs behind every table and figure of the report in both of its versions, together with the frozen reference values used to score them. It contains no analysis code and no derived tables: the report carries those, and everything in it can be recomputed from what is here. What changed in version 2. Version 1 of the report measured the external-function encoding through the solver's Python binding, which declares external functions as callbacks from a Global Interpreter Lock language and therefore evaluates them on a single worker at every requested width. Version 2 of the report replaces that entire thread ladder with runs of the same model through the C++ binding, which permits true concurrent evaluation. The two bindings are proven to evaluate identical values. This dataset version accordingly adds the 150 C++-binding ladder runs in external-function-ladder-cpp/ and keeps the superseded Python-binding ladder, shipped as external-function-ladder/ in version 1, under the explicit name external-function-ladder-python/. The slice-count study, the time-sliced ladder and the reference snapshot are unchanged. Contents. external-function-ladder-cpp/ holds 150 runs of the thread ladder on the external-function encoding through the C++ binding, which evaluates arrival times exactly on the original piecewise-linear travel-time functions: ten instances, five requested thread counts {1, 2, 4, 8, 16}, three seeds {13, 17, 42}. external-function-ladder-python/ holds the same 150 cells through the Python binding, kept as the record of what that binding delivers. slice-count-study/ holds 90 runs of the configuration study that selects the number of time slices: six instances, slice counts {5, 24, 96}, five seeds {13, 17, 42, 101, 123}, all single-threaded. time-sliced-ladder/ holds 150 runs of the thread ladder on the time-sliced encoding at 96 slices, which evaluates travel times natively from a discretized table, over the same ten instances, thread counts and seeds. Each campaign directory holds runs/, one JSON record per run carrying the configuration, the machine description, the timing, the solver's native statistics, the full incumbent trajectory with timestamps and the resource counters, and gnu-time/, the corresponding /usr/bin/time -v output recorded independently of the solver. The two external-function ladders share filenames cell for cell. The record field native.encoding distinguishes them (extfunc-cpp against extfunc). reference-precampaign.json is the frozen snapshot of best-known solutions and optimality certificates against which every gap in the report is measured. Its entry set hashes to 5e9e489169d87cde8287700501d88ebea1b60afd3731b49e17827712e34b3563. Environment. All runs executed on the grvingt cluster of the Grid'5000 testbed at the Nancy site: dual-socket Intel Xeon Gold 6130 nodes, 32 physical cores and 192 GiB of RAM per node, under Hexaly Optimizer 15.0 (build 15.0.20260724). The Python-binding ladder, the slice-count study and the time-sliced ladder ran through the solver's Python API with CPython 3.13. The C++-binding ladder declares the same external function through a compiled C++ binding. Each run was pinned with taskset to exactly as many exclusive physical cores as it requested threads. Records retain their original cluster provenance. Instances. The instances themselves are not duplicated here. They are distributed as checksummed artifacts through the MAMUT-routing benchmark platform, and each record identifies its instance by a path relative to that benchmark tree. They carry their own licenses through that platform. The bundled README.md decodes the run-tag scheme, states the reference hash and the gap conventions, and explains why some recorded incumbents are refused by the checker.