FJSPLib

The flexible jobshop scheduling problem benchmark library

FJSPLib is a comprehensive benchmark library for the Flexible Job Shop Scheduling Problem (FJSP), with three components:

Instance repository — a centralized set of classic instances (Brandimarte, Hurink, Dauzere-Perez, Chambers, Kacem, Fattahi, Behnke, Birgin)
Engine benchmark — a standardized 10-minute comparison of reference solvers (CPO, CP-SAT, OptalCP, Cplex), measuring each engine’s optimality gap and deviation from the best known bounds.
Best-known-solutions archive — a running record of the best upper/lower bounds ever found for each instance, from any source (published papers, meta-heuristics, or engine runs) — not limited to the reference engines above.

The data and source code can be found in the Github repository. This document is visible as a README.md in the Github folder flexible jobshop or as a webpage. Instances are now available in json or text formats. The standardized benchmark results for each engine on each instance is available in json format. And a json file of best known solutions is also provided with a trace of the evolution of the bounds.

Flexible jobshop instances
Standardized benchmark of engines
- A short history of the reference engines
- Comparison of reference solvers
Flexible jobshop benchmark

Overview of the flexible jobshop benchmark library

Flexible jobshop with machine-independent processing times (264)

66 x 4 instances : instances (abz, car, la, ft and orb) from the jsplib modified in 4 ways (sdata edata rdata and vdata) from Hurink, Jurisch and Thole (1994)

Flexible jobshop instances (138)

15 instances mk from Brandimarte (1993)
18 instances dpp from Dauzère-Pérès and Paulli (1994)
3 x 7 instances mt, setb4 et seti5 from Chambers and Barnes (1996)
4 instances kacem from Kacem, Hammadi and Borne (2002)
20 instances fattahi from Fattahi, Mehrabad and Jolai (2007)
60 instances behnke from Behnke and Geiger (2012)

Flexible assembly jobshop (50)

50 instances dafjs and yfjs from Birgin et al. 2014

Classification of the flexible jobshop instances

The reference engines used are

IBM ILOG Cplex : a state-of-the-art MIP engine
IBM ILOG CP Optimizer : representative of the CP-scheduling family of engines
Google CP-SAT : representative of the lazy clause generation family of engines
OptalCP : representative of the CP-scheduling family of engines

We use the following criteria to classify instances by difficulty

toy : solved to optimality (with proof) in 1 minute by at least 1 reference engine
easy : solved to optimality (with proof) in 10 minute by at least 1 reference engine
medium : solved to optimality (with proof) in 1 hour by at least 1 reference engine
hard : solved to optimality (with proof) in > 1h by at least 1 reference engine
closed : allegedly solved to optimality. Most of the time the optimal solution is known because 2 different methods independently found equal upper and lower bounds. The problem is reclassified into toy, easy, medium or hard when a reference engine is able to reproduce the results.
open : no proof of optimality

The 10 minutes time limit reflects industrial workflows in which schedules are repeatedly regenerated after manual adjustments, parameter tuning, or changes in production data. Furthermore, problems that can be solved to optimality in less than 10 minutes are suitable as sub-problems in decomposition methods like Benders decomposition, Pareto frontier generation or rolling-horizon optimization.

What we expect from a scheduling engine is:

prove optimality for as many instances as possible within the 10 minute limit
for instances that cannot be closed in 10 minutes, to minimize the deviation to the best known upper and lower bounds

Why upper and lower bounds ? Because a decomposition may use the scheduling problem as a dual certificate, not only a primal one.

Currently the instances are distributed as follows

machine-independent flexible jobshop

abz : 8 toy, 1 medium, 3 hard, 2 closed, 6 open
car : 27 toy, 3 medium, 2 closed
la : 119 toy, 17 medium, 4 hard, 11 closed, 9 open
ft : 12 toy
orb : 40 toy

flexible jobshop

mk : 9 toy, 2 medium, 2 hard, 2 open
dpp : 4 toy, 1 medium, 1 hard, 2 closed, 10 open
mt : 7 toy
setb : 7 toy
seti : 7 toy
kacem : 4 toy
fattahi : 20 toy
behnke : 15 toy, 12 medium, 1 closed, 32 open

Flexible jobshop with arbitrary precedence graphs / flexible assembly jobshop

dafjs : [status unknown]
yfjs : [status unknown]

Similar work

We have borrowed data and ideas from the following sources

Quintiq (2012 - 2014)

Around 2012, the Dutch company Quintiq started keeping track of the best known solutions for flexible jobshop problems (Barnes, Brandimarte, Dauzère and Hurink). Quintiq kept track of the lower and upper bounds, reference and date of the result. They also provided the instances and proven optimal solutions. Quintiq work is still visible thanks to the Internet Wayback Machine.

Jean-François Puget (2013)

In November 2013 Jean-François Puget benchmarked IBM ILOG CP Optimizer 12.6 on the same instances that Quintiq had worked on, improving 100 bounds and closing 69 new instances.

Time-limit	%Opt	#Improved LB	#Improved UB	#NewClosed
15mn	64.2%	48	20	52
3h	68.7%	56	35	65
24h	70.0%	59	41	69

JFP highlights the importance of clearly stated experimental protocols

A direct comparison between Quintiq Optimizer and CP Optimizer is difficult to perform because (1) the Quintiq page does not describe the experimental protocol and (2) Quintiq results are only available for the instances that were improved by Quintiq Optimizer. Nevertheless, on these instances improved by Quintiq optimizer, CP Optimizer is worse than Quintiq optimizer on 11 instances and is better on 24 instances.

Naderi, Ruiz and Roshanei (2022)

Our work was inspired by the outstanding work of Naderi, Ruiz and Roshanaei Mixed-Integer Programming versus Constraint Programming for shop scheduling problems : New Results and Outlook [NRR2022] which compares CPO, Cplex, Gurobi and OR-tools on a benchmark of 6623 instances over 17 benchmarks with a timeout of 1h. They have made all the raw results available

Scheduling Lab (2022 - present)

SchedulingLab collects instances of various types of scheduling problems, including instances not referenced here.

Dauzère-Perès, Ding, Shen and Tamssaouet and (2024)

The paper The flexible job shop scheduling problem: A review is an in-depht overview of the state-of-the-art for flexible jobshop and its variants. The supplementary material contains a summary of the best known solutions at the time of publication.

Formats

There are now three format supported in the FJSPLib

the json format
the text fjasp format
the text fjsp format

The JSON format

The JSON format follows closely the new format

For instance fattahi1 is

{
   "instance" : "fattahi1",
   "family" : "fattahi",
   "family_long" : "FattahiMehrabadJolai2007",
   "year" : "2007",
   "format" : "fajsp",
   "operations" : [
	   { "operation" : 0, "option" : 0, "machine" : 0, "duration" : 25 },
	   { "operation" : 0, "option" : 1, "machine" : 1, "duration" : 37 },
	   { "operation" : 1, "option" : 0, "machine" : 0, "duration" : 32 },
	   { "operation" : 1, "option" : 1, "machine" : 1, "duration" : 24 },
	   { "operation" : 2, "option" : 0, "machine" : 0, "duration" : 45 },
	   { "operation" : 2, "option" : 1, "machine" : 1, "duration" : 65 },
	   { "operation" : 3, "option" : 0, "machine" : 0, "duration" : 21 },
	   { "operation" : 3, "option" : 1, "machine" : 1, "duration" : 65 }
   ],
   "precedences" : [
	   { "before" : 0, "after" : 1, "job" : 0},
	   { "before" : 2, "after" : 3, "job" : 1}
   ]
}

The fjsp format

In the fjsp format

each line represents a job
the first number of the line is the number of operations in the job
then for each operation is given the number of options and as many pairs machine duration as there are options

#jobs #machines average_flexibility
#operations (#options (duration machine) (duration machine)) (#options (duration machine))

For instance fattahi1 is

2 2
2 1 25 2 37 2 1 32 2 24
2 1 45 2 65 2 1 21 2 65

meaning

(2 jobs) (2 machines) (average flexibility 2.0)
(2 operations) (2 options : (1,25) (2,37)) (2 options : (1,32) (2,24))
(2 operations) (2 options : (1,45) (2,65)) (2 options : (1,21) (2,65))

The fajsp format

The fasp format uses

the first line is the header with #operations #precedences #machines
#precedences lines of the form x y j where x and y are the indices of 2 operations, and j the number of the job that links them
#operations lines of the form #options m1 d1 m2 d2 ... mk dk where $k \in 0 .. \mathtt{#options}-1$, $m_k$ is a machine between $0$ and $\mathtt{#machines}-1$ and $d_k$ is a duration

Notice that the format doesn’t say how many jobs exist in the instance. That number needs to be deduced from the third column of the precedence data.

For instance fattahi1 is

There are 4 operations from 0 to 3 with 2 options in each operation. There are two jobs 0 -> 1 and 2 -> 3

Publications (instances)

The instances come from the following publications

Brandimarte, P (1993). Routing and scheduling in a flexible job shop by tabu search. Annals of Operations research, 41(3), 157-183.
Hurink, J., Jurisch, B., & Thole, M. (1994). Tabu search for the job-shop scheduling problem with multi-purpose machines. Operations-Research-Spektrum, 15(4), 205-215.
Dauzère-Pérès, S., & Paulli, J. (1994). Solving the general multiprocessor job-shop scheduling problem.
Chambers, J. B., & Barnes, J. W. (1996). Flexible job shop scheduling by tabu search. The University of Texas, Austin, TX, Technical Report Series ORP96-09, Graduate Program in Operations Research and Industrial Engineering.
Kacem, I., Hammadi, S., & Borne, P. (2002). Pareto-optimality approach for flexible job-shop scheduling problems: hybridization of evolutionary algorithms and fuzzy logic. Mathematics and computers in simulation, 60(3-5), 245-276.
Fattahi, P., Saidi Mehrabad, M., & Jolai, F. (2007). Mathematical modeling and heuristic approaches to flexible job shop scheduling problems. Journal of intelligent manufacturing, 18(3), 331-342.
Behnke, D., & Geiger, M. J. (2012). Test instances for the flexible job shop scheduling problem with work centers Technical report, Helmut-Schmidt-Universität, Lehrstuhl für Betriebswirtschaftslehre, insbes. Logistik-Management, RR 12-01-01.
Birgin, E. G., Feofiloff, P., Fernandes, C. G., De Melo, E. L., Oshiro, M. T., & Ronconi, D. P. (2014). A MILP model for an extended version of the flexible job shop problem. Optimization Letters, 8(4), 1417-1431.

Standardized benchmark of engines

We track the State-Of-The-Art (SOTA) of optimization engines for scheduling with a standardized 10-minute benchmark of the reference engines

The engines that are benchmarked are

IBM ILOG Cplex : a state-of-the-art MIP engine
IBM ILOG CP Optimizer : representative of the CP-scheduling family of engines
Google CP-SAT : representative of the lazy clause generation family of engines
OptalCP : representative of the CP-scheduling family of engines

A short history of the reference engines

IBM ILOG Cplex (1988 - present)

Cplex is a MIP engine founded by Robert Bixby in 1987, and acquired in 1997 by ILOG, subsequently acquired by IBM in 2009.

Like similar state-of-the-art MIP engines, Cplex features

a presolve
various LP algorithms (simplex, dual simplex, barrier)
search strategies with learning (branching, restarts)
a large collection of cuts
primal heuristics
conflict analysis adapted from SAT
solution polishing before the time-limit

Improvements in a complex software like Cplex are incremental, but we can broadly divide its evolution as follows

Cplex 1-3 [1987-1994] LP improvements (simplex, dual simplex, interior points), basic presolve and branch-and-cut. Cplex 3.0 is a performance milestone due to its mature dual simplex.
Cplex 6-7 [1998-2001] established what we would recognize today as a modern MIP (presolve + branch-and-cut + heuristics + learning) with a particular emphasis on cuts. Cplex 6.5 (1999) is considered by practitioners as a performance milestone.
Cplex 9-11 [2003-2007] more emphasis on primal approaches : heuristics (RINS, Feasibility pump), search strategies (restart + node presolve), solution polishing, SAT-like conflict analysis. Cplex 11 is considered as a performance milestone.
Cplex 11/12- [2007-] more cuts (GCD-based cuts, multi-commodity flow) sometimes mixed with heuristics (pump reduce) or search (conflict analysis), and systematic cutting plane filtering allowing a more aggressive usage of cuts

From 2008, Bixby, Rothberg and Gu created Gurobi.

References

Solving Real-World linear programs: a decade and more of progress (Robert Bixby - 2002)
Mixed Integer Programming: Analyzing 12 Years of Progress.” In Facets of Combinatorial Optimization (Tobias Achterberg, Roland Wunderling - 2013)

IBM ILOG CP Optimizer (2007 - present)

CP Optimizer is a descendant of ILOG Solver (architectured over the years by Jean-François Puget, Jean-Charles Régin and later Laurent Perron) and ILOG Scheduler (architectured by Claude Le Pape, then Philippe Laborie). CP Optimizer (led by Paul Shaw, Laurent Perron and Philippe Laborie) merged the general CP engine and the specific scheduling add-on in a single engine, promoted the model-and-run approach and pioneered a new scheduling language (optional intervals, noOverlap, cumulative functions, etc.) that has become an industry standard.

From a technical perspective CP Optimizer interleaves the following search methods

Large Neighbourhood Search (Shaw and al.) : tree-search based local search
Iterative diving (designed by Philppe Laborie) : a quick diving heuristic for “simple” scheduling problems that often provides fast and good initial solutions
Failure Directed Search (designed by Petr Vilim) : a generalization of the fail-first principle that reduces the search space by eliminating assignments unlikely to succeed
Genetic algorithms on top of the scheduling engine : named “multi-point” search, they are not on by default unless a large number of cores are available

The temporal linear relaxation solved by an LP and objective landscapes act like a reduced cost / impact based oracle but for scheduling problems.

Because CP Optimizer was designed in a time where multi-core computers weren’t common, the engine alternates the different strategies on the same core. And replicates itself over various cores with different parameters if more cores are available

The main propagation algorithms in CP optimizer are

time tabling and edge-finding for disjunctive and cumulative resources (Claude LePape, Wim Nuijten, Philippe Baptiste) later improved by Petr Vilim
logical network for implications and (dis)equalities between boolean variables
simple temporal networks for precedences

References

20+ years of scheduling with constraints at IBM/ILOG (Philippe Laborie, Jérôme Rogerie, Paul Shaw and Petr Vilim - 2018)
Introduction to CP Optimizer (Philippe Laborie - 2019)
Self-adapting large neighborhood search: Application to single-mode scheduling problems (Philippe Laborie and Daniel Godard - 2007)
Reasoning with Conditional Time-intervals (Philippe Laborie and Jérôme Rogerie - 2008)
Reasoning with Conditional Time-intervals Part II (Phillipe Laborie, Jérôme Rogerie, Paul Shaw and Petr Vilim - 2009)
Temporal linear relaxation in IBM ILOG CP Optimizer (Philippe Laborie and Jérôme Rogerie - 2014)
Failure-Directed Search for Constraint-Based Scheduling (Petr Vilim, Philippe Laborie and Paul Shaw - 2015)
Objective landscapes for constraint programming (Philippe Laborie - 2018)

Google ORTools CP-SAT (2017 - present)

CP-SAT is an open-source lazy clause generation engine augmented with an LP, MIP-style cuts and CP-style propagators designed by Laurent Perron, Frédéric Didier and Steven Gay.

CP-SAT includes

LP-based lower bounds + MIP style cuts, in particular MIP cuts specialized for scheduling
CP-style propagation algorithms (time-tabling, edge-finding)
SAT-style conflict analysis
synchronization of MIP and CP style reasonings
LNS : tree-based local search
LS with infeasible moves

CP-SAT is the “successor” of a more traditional CP + LS engine by Laurent Perron and Vincent Furnon, focusing more on VRP problems.

CP-SAT team doesn’t publish much about how CP SAT works, but maintains very informative comments in the source code. Here is an overview of the files and what they contain

Constraint	Main source files	Implemented algorithms
NoOverlap	disjunctive.cc	Detectable precedences, Edge Finding, Not-First/Not-Last, overload checking
Cumulative	cumulative.cc, timetable.cc, timetable_edgefinding.cc	Time-table propagation, Edge Finding, energetic reasoning
NoOverlap2D	diffn.cc	DiffN filtering, energetic reasoning
Circuit	circuit.cc	SCC detection, subtour elimination
AllDifferent	all_different.cc	Matching-based filtering, binary decomposition in some cases
Linear	integer_expr.cc, linear_constraint.cc	Integer propagation, pseudo-Boolean reasoning
Element	element.cc	Bounds consistency
Automaton	table.cc, automaton.cc	DFA propagation
Table	table.cc	Compact table propagation
Reservoir	reservoir.cc	Specialized cumulative reasoning

The paper From Literals to Atomic Constraints: Generalising Conflict-Driven Clause Learning for Constraint Programming contains a comparison of the implementation of various LCG-based CP solvers.

OR-Tools only creates literals for decisions. However, OR-Tools often decomposes constraints into a SAT representation, leading to more existing literals than only decision literals. During conflict analysis, an atomic constraint with no associated literal is repeatedly replaced with its reason until only existing literals are left. OR-Tools’s approach has the benefits that 1) it only creates literals that are “important” enough to be decisions, and 2) decomposed constraints ensure that there are enough literals for conflict analysis. However, OR-Tools suffers from the fact that 1) limiting created literals can lead to less general nogoods, and 2) since explanations are resolved until consisting of existing literals, explanation lifting and nogood minimisation can have less impact

The presolver of CP-SAT does a significant amount of work, closer to a MIP than a typical CP engine

variable fixing
domain tightening
affine relation detection
duplicate constraint elimination
implied bound computation
linear simplification
clique extraction
symmetry detection
interval simplification
objective simplification

Moreover some of the classic scheduling algorithms have been transformed into cutting planes (scheduling_cuts.cc) : energetic reasoning, time-table propagation, cumulative precedence, cumulative completion time, energetic reasoning for disjunctive resources.

References

CP-SAT at scheduling seminar (Laurent Perron - 2024)
From Literals to Atomic Constraints: Generalising Conflict-Driven Clause Learning for Constraint Programming (Imko Marijnissen, Maarten Flippo, Emir Demirović - CP2026)

OptalCP (2021 - present)

OptalCP was architectured by Petr Vilim, Nicolas Bonifas and Diego Olivier Fernandez Pons (initially with input from Philippe Laborie). Compared to CPO the parallelism is done with one strategy per core instead of interleaving. The strategies used are

Large Neighbourhood Search (LNS) : tree-search based local search
Failure Directed Search (FDS) : generalizes first-fail principle
FDSDual : generalizes destructive lower bounds

OptalCP continues the legacy of CP Optimizer (engine style, modeling language). The hybridization of OptalCP with heuristics and meta-heuristics is done outside by communicating upper and lower bounds in real time (during search)

References

OptalCP at scheduling seminar (Petr Vilim - 2026)

If your name appears in this section and you notice an error contact me

Comparison of reference solvers

We adopt the following metrics to compare the engines

Geometric average of lower bound to best known lower bound ratio

\[LB_{avg} = \exp\left(\sum_k\log\frac{LB}{LB_{best}}\right)\]

Geometric average of the upper bound to best known upper bound ratio

\[UB_{avg} = \exp\left(\sum_k\log\frac{UB}{UB_{best}}\right)\]

Geometric shifted average of the gap

\[GAP = \exp\left(\sum_k\log\left(1 + \frac{UB - LB}{UB}\right)\right) - 1\]

Comparisons were performed on a Windows PC with an i7 4-core 3GHz 32GB ram in 600 seconds

Cplex 22.1.1.0
CPO 22.1.1.0
- with gap tolerance = 0
CP-SAT V9.15.6755 with default configuration
OptalCP Academic Version 2026.4.0
- with maximum propagation instead of default, gap tolerance = 0 and some other parameter changes
- we may benchmark with default parameters later

The raw data is in the solutions folder

We recommend to run your own benchmarks on your own machines. All required code is provided with HOWTO instructions in each README file.

Important caveats

Engines have relative and absolute optimality tolerances (CPO, OptalCP) which can lead to reporting sub-optimal solutions as optimal (for the tolerance). For this test the tolerances have been set to zero
Engines are very non-deterministic (results between two runs of the same engine on the same machine differ significantly) due to parallelism. As a result it makes no sense to consider very accurate values for average deviations or gap.
Engines have different bottlenecks (CPU, memory) that are due to their internal architecture and trade-offs made by their designers. An engine doesn’t behave in the same way with 2, 4, 8, 16, 32 or 64 cores, doesn’t behave the same in machines with fast / slow memory, etc.
Engines recommend different configurations : CP-SAT recommends a minimum of 8 cores and ideally 16 (due to the high number of strategies that need to be interleaved with less cores). CPO on the other hand gives most of its performance on 1 or 2 cores . The 4 core configuration was chosen because it is a usual laptop configuration, but commercial users will probably run on a larger server.

Best known solutions

In this section are collected the best known solutions (upper and lower bound) for each problem in the benchmark.

The solutions may come from

Published papers (eg. MG2000), the section publications provides references
An engine run by someone else (eg. CPO2013) which results have been published
An engine run by us with approximate resolution time

The type of hardware and time required to find the best known solution are difficult to track and compare, in particular for bounds coming from published papers. Which is why

When a reference engine reproduces a published bound, the table credits the engine because of the reproducibility advantage
An approximative timing for reference engines is provided, in particular when the time to find the solution is unusually long

We do not systematically run the instances for very long times on large machines. Most of the instances that appear as having been solved after a large comptation time (eg. 40h) had peculiarities (e.g. best lb + 1 == best ub) that justified exploring how long it would take to solve them to optimality. We also devote more effort to solve instances which best known solutions are given by papers that are old, difficult to find and difficlt to reproduce. This allows verifying the paper claims and having a more accessible way of generating the result.

Best known solutions json format

The best known solutions are now collected in a json file with the following syntax

{
    "instance":"dpp06a",
    "size":"10 x 5",
    "type":"flexible jobshop",
    "family":"dpp",
    "family_long":"Dauzère-Pérès and Paulli (1994)",
    "status":"open",
    "lower_bound":2164,
    "upper_bound":2169,
    "lb_data":[{
        "value":2164,
        "date":"2026-01-01",
        "solver":"CdGKGC2025",
        "time":"?",
        "hardware":"?",
        "certificate":"no"
    }],
    "ub_data":[{
        "value":2169,
        "date":"2026-01-01",
        "solver":"CdGKGC2025",
        "time":"?",
        "hardware":"?",
        "certificate":"no"
    }]
}

For most of the best known solutions, the date, hardware, running time and certificate (valid primal or valid dual solution) are not known. The data will be progressively updated to the best of our knowledge.

Best known solutions per instance family

Hurink, Jurisch and Thole (1994) - Machine independent processing times

The problems in this benchmark are modified versions of the corresponding jobshop problems. They are divided into

sdata : each operation can be assigned to a single machine (jobshop)
edata : a few operations can be assigned to many machines
rdata : most operations can be assigned to a few machines
vdata : all operation can be assigned to many machines

Instance	Size	sdata	edata	rdata	vdata	Solved by
abz5	10 x 10	1234	1167	954	859	OptalCP
abz6	10 x 10	943	925	807	742	OptalCP
abz7	20 x 15	656	604 / 610	497 / 522	492	OptalCP \| CPO2013 / Quintiq \| CdGKGC2025 / DLLSXG2019 \| Quintiq
abz8	20 x 15	667	625 / 636	509 / 535	506 / 507	OptalCP \| Quintiq / CPO2013 \| CdGKGC2025 / DLLSXG2019 \| OptalCP / Quintiq
abz9	20 x 15	678	644	517 / 536	497	OptalCP \| CPO2013 \| CPO2013 / Quintiq \| OptalCP / Quintiq

Instance	Size	sdata	edata	rdata	vdata	Solved by
car1	11 x 5	7038	6176	5034	5005	OptalCP
car2	13 x 4	7166	6327	5985	5929	OptalCP
car3	12 x 5	7312	6856	5622	5597	OptalCP \| OptalCP / Quintiq \| OptalCP
car4	14 x 4	8003	7789	6514	6514	OptalCP
car5	10 x 6	7702	7229	5615	4909	OptalCP \| OptalCP / CdGKGC2025
car6	8 x 9	8313	7990	6147	5486	OptalCP
car7	7 x 7	6558	6123	4425	4281	OptalCP
car8	8 x 8	8264	7689	5692	4613	OptalCP

Instance	Size	sdata	edata	rdata	vdata	Solved by
la01	10 x 5	666	609	570	570	OptalCP
la02	10 x 5	655	655	529	529	OptalCP
la03	10 x 5	597	550	477	477	OptalCP
la04	10 x 5	590	568	502	502	OptalCP
la05	10 x 5	593	503	457	457	OptalCP
la06	15 x 5	926	833	799	799	OptalCP
la07	15 x 5	890	762	749	749	OptalCP
la08	15 x 5	863	845	765	765	OptalCP
la09	15 x 5	951	878	853	853	OptalCP
la10	15 x 5	958	866	804	804	OptalCP
la11	20 x 5	1222	1103	1071	1071	OptalCP
la12	20 x 5	1039	960	936	936	OptalCP
la13	20 x 5	1150	1053	1038	1038	OptalCP
la14	20 x 5	1292	1123	1070	1070	OptalCP
la15	20 x 5	1207	1111	1089	1089	OptalCP
la16	10 x 10	945	892	717	717	OptalCP
la17	10 x 10	784	707	646	646	OptalCP
la18	10 x 10	848	842	666	663	OptalCP
la19	10 x 10	842	796	700	617	OptalCP
la20	10 x 10	902	857	756	756	OptalCP
la21	15 x 10	1046	1009	809 / 825	800	OptalCP \| CdGKGC2025 / Quintiq \| OptalCP
la22	15 x 10	927	880	745 / 753	733	OptalCP \| CdGKGC2025 / DLLSXG2019 \| OptalCP / CPO2013
la23	15 x 10	1032	950	820 / 831	809	OptalCP \| CdGKGC2025 / DLLSXG2019 \| OptalCP
la24	15 x 10	935	908	780 / 795	773	OptalCP \| CdGKGC2025 / DLLSXG2019 \| OptalCP
la25	15 x 10	977	936	771 / 779	751	OptalCP \| CdGKGC2025 / DLLSXG2019 \| OptalCP / Quintiq
la26	20 x 10	1218	1106	1056 / 1057	1052	OptalCP \| DOFP2026a / Quintiq \| OptalCP
la27	20 x 10	1235	1181	1085	1084	OptalCP \| DOFP2026a / Quintiq \| OptalCP
la28	20 x 10	1216	1142	1075 / 1076	1069	OptalCP \| CPO2013 \| DOFP2026a / Quintiq \| OptalCP
la29	20 x 10	1152	1107	993 / 994	993	OptalCP \| CPO2013 \| OptalCP / Quintiq
la30	20 x 10	1355	1188	1068 / 1071	1068	OptalCP \| CPO2013 \| OptalCP / Quintiq
la31	30 x 10	1784	1532	1520	1520	OptalCP
la32	30 x 10	1850	1698	1657	1657	OptalCP \| OptalCP / Quintiq
la33	30 x 10	1719	1547	1497	1497	OptalCP \| OptalCP / Quintiq \| OptalCP / MG2000
la34	30 x 10	1721	1599	1535	1535	OptalCP \| OptalCP / Quintiq \| OptalCP
la35	30 x 10	1888	1736	1549	1549	OptalCP \| OptalCP / Quintiq \| OptalCP
la36	15 x 15	1268	1160	1023	948	OptalCP
la37	15 x 15	1397	1397	1062	986	OptalCP
la38	15 x 15	1196	1141	954	943	OptalCP
la39	15 x 15	1233	1184	1011	922	OptalCP
la40	15 x 15	1222	1144	955	955	OptalCP

Instance	Size	sdata	edata	rdata	vdata	Solved by
ft06	6 x 6	55	55	47	47	OptalCP
ft10	10 x 10	930	871	686	655	OptalCP
ft20	20 x 5	1165	1088	1022	1022	OptalCP

Instance	Size	sdata	edata	rdata	vdata	Solved by
orb1	10 x 10	1059	977	746	695	OptalCP
orb2	10 x 10	888	865	696	620	OptalCP
orb3	10 x 10	1005	951	712	648	OptalCP
orb4	10 x 10	1005	984	753	753	OptalCP
orb5	10 x 10	887	842	639	584	OptalCP
orb6	10 x 10	1010	958	754	715	OptalCP
orb7	10 x 10	397	389	302	275	OptalCP
orb8	10 x 10	899	894	639	573	OptalCP
orb9	10 x 10	934	933	694	659	OptalCP
orb10	10 x 10	944	933	742	681	OptalCP

FT instances are also known as MT because the 1963 paper of Fisher and Thompson was published in the book “Industrial scheduling” by Muth and Thompson.

Brandimarte (1993)

Instance	Size	Problem	LB	UB	Type	Solved by
mk01	10 x 6	flexible jobshop	40	40	toy	OptalCP in < 1m
mk02	10 x 6	flexible jobshop	26	26	toy	OptalCP in < 1m
mk03	15 x 8	flexible jobshop	204	204	toy	OptalCP in < 1m
mk04	15 x 8	flexible jobshop	60	60	toy	OptalCP in < 1m
mk05	15 x 4	flexible jobshop	172	172	toy	OptalCP in < 1m
mk06	10 x 15	flexible jobshop	57	57	medium	OptalCP in < 1h
mk07	20 x 5	flexible jobshop	139	139	medium	OptalCP in < 1h
mk08	20 x 10	flexible jobshop	523	523	toy	OptalCP in < 1m
mk09	20 x 10	flexible jobshop	307	307	toy	OptalCP in < 1m
mk10	20 x 15	flexible jobshop	189	193	open	lb DOFP2026a \| ub Quintiq
mk11	30 x 5	flexible jobshop	609	609	hard	OptalCP in < 24h
mk12	30 x 10	flexible jobshop	508	508	toy	OptalCP in < 1m
mk13	30 x 10	flexible jobshop	382	390	open	lb DOFP2026a \| ub OptalCP
mk14	30 x 15	flexible jobshop	694	694	toy	OptalCP in < 1m
mk15	30 x 15	flexible jobshop	333	333	hard	OptalCP in < 2h

Instances mk11 to mk15 are present in the supplementary material of Test Instances for the Flexible Job Shop Scheduling Problem with Work Centers but absent from other problem repositories

Dauzère-Pérès and Paulli (1994)

Instance	Size	Problem	LB	UB	Type	Solved by
dpp01a	10 x 5	flexible jobshop	2505	2505	toy	OptalCP in < 1m
dpp02a	10 x 5	flexible jobshop	2228	2228	toy	OptalCP in < 1m
dpp03a	10 x 5	flexible jobshop	2228	2228	toy	OptalCP in < 1m
dpp04a	10 x 5	flexible jobshop	2503	2503	toy	OptalCP in < 1m
dpp05a	10 x 5	flexible jobshop	2195	2199	open	CdGKGC2025
dpp06a	10 x 5	flexible jobshop	2164	2169	open	CdGKGC2025
dpp07a	15 x 8	flexible jobshop	2216	2254	open	lb CPO2013 \| ub DLLSXG2019
dpp08a	15 x 8	flexible jobshop	2061	2061	hard	OptalCP in < 10h
dpp09a	15 x 8	flexible jobshop	2061	2061	medium	OptalCP in < 1h
dpp10a	15 x 8	flexible jobshop	2212	2241	open	lb CPO2013 \| ub Quintiq
dpp11a	15 x 8	flexible jobshop	2019	2037	open	lb CdGKGC2025 \| ub Quintiq
dpp12a	15 x 8	flexible jobshop	1969	1984	open	lb OptalCP \| ub Quintiq
dpp13a	20 x 10	flexible jobshop	2206	2236	open	lb CdGKGC2025 \| ub DLLSXG2019
dpp14a	20 x 10	flexible jobshop	2161	2161	closed	lb OptalCP \| ub Quintiq
dpp15a	20 x 10	flexible jobshop	2161	2161	closed	lb OptalCP \| ub Quintiq
dpp16a	20 x 10	flexible jobshop	2202	2231	open	lb CdGKGC2025 \| ub Quintiq
dpp17a	20 x 10	flexible jobshop	2089	2105	open	lb CdGKGC2025 \| ub Quintiq
dpp18a	20 x 10	flexible jobshop	2057	2070	open	lb OptalCP \| ub Quintiq

Chambers and Barnes (1996)

Instance	Size	Problem	LB	UB	Type	Solved by
mt10c1	10 x 11	flexible jobshop	927	927	toy	OptalCP in < 1m
mt10cc	10 x 12	flexible jobshop	908	908	toy	OptalCP in < 1m
mt10x	10 x 11	flexible jobshop	918	918	toy	OptalCP in < 1m
mt10xx	10 x 12	flexible jobshop	918	918	toy	OptalCP in < 1m
mt10xxx	10 x 13	flexible jobshop	918	918	toy	OptalCP in < 1m
mt10xy	10 x 12	flexible jobshop	905	905	toy	OptalCP in < 1m
mt10xyz	10 x 13	flexible jobshop	847	847	toy	OptalCP in < 1m
setb4c9	15 x 11	flexible jobshop	914	914	toy	OptalCP in < 1m
setb4cc	15 x 12	flexible jobshop	907	907	toy	OptalCP in < 1m
setb4x	15 x 11	flexible jobshop	925	925	toy	OptalCP in < 1m
setb4xx	15 x 12	flexible jobshop	925	925	toy	OptalCP in < 1m
setb4xxx	15 x 13	flexible jobshop	925	925	toy	OptalCP in < 1m
setb4xy	15 x 12	flexible jobshop	910	910	toy	OptalCP in < 1m
setb4xyz	15 x 13	flexible jobshop	902	902	toy	OptalCP in < 1m
seti5c12	15 x 16	flexible jobshop	1169	1169	toy	OptalCP in < 1m
seti5cc	15 x 17	flexible jobshop	1135	1135	toy	OptalCP in < 1m
seti5x	15 x 16	flexible jobshop	1198	1198	toy	OptalCP in < 1m
seti5xx	15 x 17	flexible jobshop	1194	1194	toy	OptalCP in < 1m
seti5xxx	15 x 18	flexible jobshop	1194	1194	toy	OptalCP in < 1m
seti5xy	15 x 17	flexible jobshop	1135	1135	toy	OptalCP in < 1m
seti5xyz	15 x 18	flexible jobshop	1125	1125	toy	OptalCP in < 1m

Kacem, Hammadi and Borne (2002)

Instance	Size	Problem	LB	UB	Type	Solved by
kacem1	4 x 6	flexible jobshop	11	11	toy	OptalCP in < 1m
kacem2	10 x 7	flexible jobshop	11	11	toy	OptalCP in < 1m
kacem3	10 x 10	flexible jobshop	7	7	toy	OptalCP in < 1m
kacem4	15 x 10	flexible jobshop	11	11	toy	OptalCP in < 1m

Fattahi, Mehrabad and Jolai (2007)

Instance	Size	Problem	LB	UB	Type	Solved by
fattahi1	2 x 2	flexible jobshop	66	66	toy	OptalCP in < 1m
fattahi2	2 x 2	flexible jobshop	107	107	toy	OptalCP in < 1m
fattahi3	3 x 2	flexible jobshop	221	221	toy	OptalCP in < 1m
fattahi4	3 x 2	flexible jobshop	355	355	toy	OptalCP in < 1m
fattahi5	3 x 2	flexible jobshop	119	119	toy	OptalCP in < 1m
fattahi6	3 x 2	flexible jobshop	320	320	toy	OptalCP in < 1m
fattahi7	3 x 5	flexible jobshop	397	397	toy	OptalCP in < 1m
fattahi8	3 x 4	flexible jobshop	253	253	toy	OptalCP in < 1m
fattahi9	3 x 3	flexible jobshop	210	210	toy	OptalCP in < 1m
fattahi10	4 x 5	flexible jobshop	516	516	toy	OptalCP in < 1m
fattahi11	5 x 6	flexible jobshop	468	468	toy	OptalCP in < 1m
fattahi12	5 x 7	flexible jobshop	446	446	toy	OptalCP in < 1m
fattahi13	6 x 7	flexible jobshop	466	466	toy	OptalCP in < 1m
fattahi14	7 x 7	flexible jobshop	554	554	toy	OptalCP in < 1m
fattahi15	7 x 7	flexible jobshop	514	514	toy	OptalCP in < 1m
fattahi16	8 x 7	flexible jobshop	634	634	toy	OptalCP in < 1m
fattahi17	8 x 7	flexible jobshop	879	879	toy	OptalCP in < 1m
fattahi18	9 x 8	flexible jobshop	884	884	toy	OptalCP in < 1m
fattahi19	11 x 8	flexible jobshop	1055	1055	toy	OptalCP in < 1m
fattahi20	12 x 8	flexible jobshop	1196	1196	toy	OptalCP in < 1m

Behnke and Geiger (2012)

Instance	Size	Problem	LB	UB	Type	Solved by
behnke1	10 x 20	flexible jobshop	90	90	toy	OptalCP in < 1m
behnke2	10 x 20	flexible jobshop	91	91	toy	OptalCP in < 1m
behnke3	10 x 20	flexible jobshop	91	91	toy	OptalCP in < 1m
behnke4	10 x 20	flexible jobshop	97	97	toy	OptalCP in < 1m
behnke5	10 x 20	flexible jobshop	91	91	toy	OptalCP in < 1m
behnke6	20 x 20	flexible jobshop	125	125	medium	OptalCP in < 1h
behnke7	20 x 20	flexible jobshop	117	124	open	OptalCP
behnke8	20 x 20	flexible jobshop	123	123	medium	OptalCP in < 1h
behnke9	20 x 20	flexible jobshop	125	125	medium	OptalCP in < 1h
behnke10	20 x 20	flexible jobshop	127	127	medium	OptalCP in < 1h
behnke11	50 x 20	flexible jobshop	223	228	open	lb DOFP2026a \| ub OptalCP
behnke12	50 x 20	flexible jobshop	213	219	open	lb DOFP2026a \| ub OptalCP
behnke13	50 x 20	flexible jobshop	223	229	open	lb DOFP2026a \| ub OptalCP
behnke14	50 x 20	flexible jobshop	221	230	open	lb DOFP2026a \| ub OptalCP
behnke15	50 x 20	flexible jobshop	219	228	open	lb DOFP2026a \| ub OptalCP
behnke16	100 x 20	flexible jobshop	391	412	open	lb DOFP2026a \| ub OptalCP
behnke17	100 x 20	flexible jobshop	392	401	open	lb DOFP2026a \| ub OptalCP
behnke18	100 x 20	flexible jobshop	390	396	open	lb DOFP2026a \| ub OptalCP
behnke19	100 x 20	flexible jobshop	395	400	open	lb DOFP2026a \| ub OptalCP
behnke20	100 x 20	flexible jobshop	391	398	open	lb DOFP2026a \| ub OptalCP
behnke21	10 x 40	flexible jobshop	85	85	toy	OptalCP in < 1m
behnke22	10 x 40	flexible jobshop	87	87	toy	OptalCP in < 1m
behnke23	10 x 40	flexible jobshop	85	85	toy	OptalCP in < 1m
behnke24	10 x 40	flexible jobshop	87	87	toy	OptalCP in < 1m
behnke25	10 x 40	flexible jobshop	87	87	toy	OptalCP in < 1m
behnke26	20 x 40	flexible jobshop	113	113	medium	OptalCP in < 1h
behnke27	20 x 40	flexible jobshop	122	122	medium	OptalCP in < 1h
behnke28	20 x 40	flexible jobshop	114	114	medium	OptalCP in < 1h
behnke29	20 x 40	flexible jobshop	116	117	open	lb DOFP2026a \| ub OptalCP
behnke30	20 x 40	flexible jobshop	120	120	medium	OptalCP in < 1h
behnke31	50 x 40	flexible jobshop	226	226	closed	lb DOFP2026a \| ub OptalCP
behnke32	50 x 40	flexible jobshop	220	224	open	lb DOFP2026a \| ub OptalCP
behnke33	50 x 40	flexible jobshop	223	224	open	lb DOFP2026a \| ub OptalCP
behnke34	50 x 40	flexible jobshop	219	223	open	lb DOFP2026a \| ub OptalCP
behnke35	50 x 40	flexible jobshop	211	214	open	lb DOFP2026a \| ub OptalCP
behnke36	100 x 40	flexible jobshop	381	388	open	lb DOFP2026a \| ub OptalCP
behnke37	100 x 40	flexible jobshop	387	391	open	lb DOFP2026a \| ub OptalCP
behnke38	100 x 40	flexible jobshop	386	389	open	lb DOFP2026a \| ub OptalCP
behnke39	100 x 40	flexible jobshop	384	389	open	lb DOFP2026a \| ub OptalCP
behnke40	100 x 40	flexible jobshop	415	419	open	OptalCP
behnke41	10 x 60	flexible jobshop	87	87	toy	OptalCP in < 1m
behnke42	10 x 60	flexible jobshop	87	87	toy	OptalCP in < 1m
behnke43	10 x 60	flexible jobshop	86	86	toy	OptalCP in < 1m
behnke44	10 x 60	flexible jobshop	84	84	toy	OptalCP in < 1m
behnke45	10 x 60	flexible jobshop	87	87	toy	OptalCP in < 1m
behnke46	20 x 60	flexible jobshop	114	114	medium	OptalCP in < 1h
behnke47	20 x 60	flexible jobshop	117	117	medium	OptalCP in < 1h
behnke48	20 x 60	flexible jobshop	124	125	open	lb DOFP2026a \| ub OptalCP
behnke49	20 x 60	flexible jobshop	113	113	medium	OptalCP in < 1h
behnke50	20 x 60	flexible jobshop	123	123	medium	OptalCP in < 1h
behnke51	50 x 60	flexible jobshop	215	218	open	lb DOFP2026a \| ub OptalCP
behnke52	50 x 60	flexible jobshop	210	212	open	lb DOFP2026a \| ub OptalCP
behnke53	50 x 60	flexible jobshop	211	215	open	lb DOFP2026a \| ub OptalCP
behnke54	50 x 60	flexible jobshop	221	223	open	lb DOFP2026a \| ub OptalCP
behnke55	50 x 60	flexible jobshop	221	223	open	lb DOFP2026a \| ub OptalCP
behnke56	100 x 60	flexible jobshop	384	390	open	lb DOFP2026a \| ub OptalCP
behnke57	100 x 60	flexible jobshop	385	390	open	lb DOFP2026a \| ub OptalCP
behnke58	100 x 60	flexible jobshop	392	397	open	lb DOFP2026a \| ub OptalCP
behnke59	100 x 60	flexible jobshop	392	398	open	lb DOFP2026a \| ub OptalCP
behnke60	100 x 60	flexible jobshop	397	402	open	lb DOFP2026a \| ub OptalCP

Birgin et al. (2014) - arbitrary precedence DAGs

[Results to be published soon]

Publications (best known solutions)

The upper and lower bounds come from

MG2000 (1 bounds in la) : M. Mastrolilli, L. Gambardella, Effective neighbourhood functions for the flexible job shop problem, Journal of Scheduling 3 (2000) 3–20
Quintiq (31 bounds in mk, dpp, abz, car and la) : Quintiq http://www.quintiq.com/optimization/fjssp-world-records.html (2013) - this site doesn’t exist anymore
CPO2013 (14 bounds in abz, la and dpp) : Jean-François Puget Solving flexible job shop scheduling problems (cp optimizer 12.6) https://www.ibm.com/developerworks/community/blogs/jfp/entry/solving_flexible_job_shop_scheduling_problems?lang=en (2013) - this site doesn’t exist anymore
DLLSXG2019 (8 bounds in dpp, abz and la) : J. Ding, Z. Lu, C.-M. Li, L. Shen, L. Xu, F. Glover (2019) A two-individual based evolutionary algorithm for the flexible job shop scheduling problem, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33, 2019, pp. 2262–2271
CdGKGC2025 (14 bounds in #a, abz, car, la) : Marc-Emmanuel Coupvent des Graviers, Lotfi Kobrosly, Christophe Guettier, and Tristan Cazenave (2025). Updating Lower and Upper Bounds for the Job-Shop Scheduling Problem Test Instances CoRR abs/2504.16106
DOFP2026a (36 bounds in mk, behnke and la) : Diego Olivier Fernandez Pons (2026) Personal communication, study / test to improve OptalCP

All other bounds were found with OptalCP

FJSPLib

Scheduling benchmarks for optimization engines

The flexible jobshop scheduling problem benchmark library

Table of Contents

Overview of the flexible jobshop benchmark library

Classification of the flexible jobshop instances

Similar work

Quintiq (2012 - 2014)

Jean-François Puget (2013)

Naderi, Ruiz and Roshanei (2022)

Scheduling Lab (2022 - present)

Dauzère-Perès, Ding, Shen and Tamssaouet and (2024)

Formats

The JSON format

The fjsp format

The fajsp format

Publications (instances)

Standardized benchmark of engines

A short history of the reference engines

IBM ILOG Cplex (1988 - present)

IBM ILOG CP Optimizer (2007 - present)

Google ORTools CP-SAT (2017 - present)

OptalCP (2021 - present)

Comparison of reference solvers

Best known solutions

Best known solutions json format

Best known solutions per instance family

Hurink, Jurisch and Thole (1994) - Machine independent processing times

Brandimarte (1993)

Dauzère-Pérès and Paulli (1994)

Chambers and Barnes (1996)

Kacem, Hammadi and Borne (2002)

Fattahi, Mehrabad and Jolai (2007)

Behnke and Geiger (2012)

Birgin et al. (2014) - arbitrary precedence DAGs

Publications (best known solutions)