Dimacs2021 (VRPTW)¶
Dimacs2021 is the MAMUT-routing curation of the VRPTW convention introduced by the DIMACS VRPTW competition during the 12th DIMACS Implementation Challenge. It reuses the same Solomon and Gehring--Homberger instance universe as Sintef2008, but changes the evaluation contract in two crucial ways: costs are integerized, and the objective is single-objective total-cost minimization, subject to the instance vehicle limit.
This change was motivated by long-standing difficulties with the classical SINTEF convention. Floating point arc costs make exact comparison delicate because of rounding errors and non-associativity. The hierarchical objective also complicates exact methods: minimizing fleet size first and distance second can require a two-stage optimization strategy, or at least solver-specific handling of lexicographic objectives. For many exact-method papers, a pure cost objective was therefore more natural, but for years the community lacked a widely accepted cost-only counterpart to SINTEF.
The DIMACS rules proposed such a counterpart. Euclidean distances are scaled by a factor 10 and truncated to integers. A DIMACS cost such as 8273 is therefore comparable, but not identical, to a SINTEF-style floating point distance around 827.3. This integer contract makes objective values reproducible across languages and solvers, and it avoids many false improvements caused by reporting precision.
The DIMACS competition had a lasting impact because it gave heuristic and exact-method teams a shared, checker-oriented target. It also influenced modern solver tooling: for example, PyVRP, built around Hybrid Genetic Search, naturally supports the DIMACS-style cost-minimization convention.
The original DIMACS benchmark material, however, was not distributed as a complete curated BKS repository. Public result tables were mainly provided through scoreboards and spreadsheets without route files. The MAMUT-routing Dimacs2021 family fills this gap by collecting, validating, and normalizing route-level BKS files under the DIMACS objective. Several BKS sources were used, including CVRPLib-provided Solomon/Homberger solutions reevaluated under the DIMACS contract, PyVRP instance collections for larger cases, and additional personal experiments run on Grid5000.
Licensing note: MAMUT-routing-authored curation artifacts for this family are distributed under the MIT License where MAMUT-routing holds the relevant rights. The underlying historical benchmark definitions, competition material, and some BKS sources remain third-party material and are not relicensed by this curation.
The family currently mirrors the 468 classical instances over 8 different instance sizes, each with a MonoCost BKS.
July 2026 note. CVRPLib's newer VRPTW instances and BKS follow the DIMACS convention (scaled, integerized arc costs with mono-cost minimization), as clarified by Leon Lan, one of the PyVRP maintainers. Their BKS are therefore directly comparable with Dimacs2021 whenever the remaining conventions (rounding/truncation and fleet limits) also match. An earlier note here incorrectly stated that CVRPLib's VRPTW material used float64 arc costs.
Data
Instances and best-known solutions: /benchmarks/vrptw/dimacs2021/ on the website. Repository: https://github.com/ANR-MAMUT/MAMUT-routing/tree/main/benchmarks/VRPTW/Dimacs2021.
Repository README¶
Dimacs2021 benchmark family¶
This directory contains the MAMUT-routing curation of the classical Solomon/Gehring-Homberger VRPTW instance universe under the DIMACS 2021 integer cost-minimization convention.
License scope:
- MAMUT-routing-authored curation artifacts, normalized metadata, and contributed files are distributed under MIT where MAMUT-routing holds the relevant rights.
- The underlying instance universe and some BKS sources originate from third-party historical benchmark and competition material. This directory-level notice does not relicense those third-party sources.
Preserve this README, this directory's LICENSE file, and the repository NOTICE when redistributing this benchmark family.