https://scheduleopt.github.io/benchmarks/
TLDR
Just a dataset of the most known scheduling problems, with some reference engines (in case we wanted to try solving them) and with the best known LB and UB of each instance collected from differetn sources.
The site defines the library as three parts:
- instances An instance repository. It holds benchmark instances from the literature in one shared JSON format.
- engine run An engine benchmark. Four reference engines run each instance for 10 minutes on the same machine.
- archive A best-known-solutions archive. It keeps the best lower and upper bound for each instance, with the history of who found each one.
2 families of problems
- JSPLib: job-shop scheduling. A job is a fixed sequence of operations, and each operation must use one given machine. A machine runs one operation at a time.
- FJSPLib: flexible job-shop scheduling. Each operation can use one machine from a set of eligible machines. Each option has its own duration. Thus, the solver must assign machines and also order the operations.
In more detail:
Figure from ScheduleOpt scheduling benchmark, scheduleopt.github.io/benchmarks (read 2026-10-17). Open full size
REALM-Bench uses some of these problems as well like (DMU, TA, ABZ, SWV, YN), J1 results numbers can be replaced directly with the results in this project.
Difficulty classes
The archive puts each instance in one class. The class depends on the time that the fastest reference engine needs to prove optimality— historicallyy:
- toy: proof in 1 minute. easy: proof in 10 minutes. medium: proof in 1 hour. hard: proof in more than 1 hour.
- closed: the optimum is known, usually because two independent methods found equal bounds. open: no proof of optimality.
For example JSPLib problems results:
Figure from ScheduleOpt scheduling benchmark, scheduleopt.github.io/benchmarks (read 2026-10-17). Open full size