1: \documentclass[11pt]{article}

2: 

3: % Make \epsilon and \varepsilon consistently \varepsilon

4: \let\epsilon=\varepsilon

5: 

6: \usepackage{fullpage}

7: \usepackage{epsfig}

8: \usepackage{amssymb}

9: 

10: \newtheorem{thm}{Theorem}

11: \newtheorem{lem}[thm]{Lemma}

12: \newtheorem{conj}[thm]{Conjecture}

13: \newtheorem{cor}[thm]{Corollary}

14: \newtheorem{dfn}[thm]{Definition}

15: \newtheorem{fact}[thm]{Fact}

16: \newtheorem{claim}[thm]{Claim}

17: 

18: \newenvironment{pf}{{\bf Proof:}}{\hfill $\Box$ \bigbreak}

19: 

20: \begin{document}

21: 

22: \title{The Lazy Bureaucrat Scheduling Problem\thanks{

23: Appears in preliminary form in

24: {\em Proceedings of the Workshop on Discrete Algorithms (WADS)},

25: Vancouver, BC, Canada, Aug 11--14, 1999, pages 

26: 122--133~\cite{ArkinBenderMitchellSkiena99}.}}

27: 

28: \author{Esther M. Arkin\thanks{estie@ams.sunysb.edu. 

29: Department of Applied Mathematics and Statistics.

30: State University of New York, Stony Brook, NY 11794-3600.

31: Partially supported by 

32: the National Science Foundation (CCR-9732220, CCR-0098172) and

33: grants from ISX Corp.\ and HRL Laboratories.}

34: \and

35: Michael A. Bender\thanks{bender@cs.sunysb.edu.

36: Department of Computer Science.

37: State University of New York, Stony Brook, NY 11794-4400.

38: Partially supported by 

39: grants from ISX Corp., HRL Laboratories, Sandia National 

40: Labs, and the National Science Foundation (EIA-0112849, CCR-0208670).}

41: %

42: \and

43: Joseph S. B. Mitchell\thanks{jsbm@ams.sunysb.edu.

44: Department of Applied Mathematics and Statistics.

45: State University of New York, Stony Brook, NY 11794-3600.

46: Supported in part by 

47: HRL Laboratories, Honda Fundamental Research Labs, ISX Corp., Metron Aviation, Inc., Sandia National 

48: Labs, Seagull Technologies, Sun Microsystems, and 

49: the National Science Foundation (CCR-9732220, CCR-0098172).}

50: %

51: \and

52: Steven S. Skiena\thanks{skiena@cs.sunysb.edu.

53: Department of Computer Science.

54: State University of New York, Stony Brook, NY 11794-4400.

55: Supported in part by

56: the National Science Foundation (CCR-9988112) and

57: by the Office of Naval Research (N00149710589).}

58: }

59: 

60: \maketitle

61: 

62: \begin{abstract}

63:   We introduce a new class of scheduling problems in which the

64:   optimization is performed by the worker (single ``machine'') who

65:   performs the tasks.  A typical worker's objective is to minimize the

66:   amount of work he does (he is ``lazy''), or more generally, to

67:   schedule as inefficiently (in some sense) as possible.  The worker

68:   is subject to the constraint that he must be busy when there is work

69:   that he {\em can\/} do; we make this notion precise both in the

70:   preemptive and nonpreemptive settings.  The resulting class of

71:   ``perverse'' scheduling problems, which we denote ``Lazy Bureaucrat

72:   Problems,'' gives rise to a rich set of new questions that explore

73:   the distinction between maximization and minimization in computing

74:   optimal schedules.

75: \end{abstract}

76: 

77: \noindent {\bf Keywords:}\quad

78: Scheduling, Approximation Algorithms,

79: Optimization, Dynamic Programming, NP-completeness, Lazy Bureaucrat.

80: 68M20, 68Q25, 90B35, 90B70.

81: 

82: \section{Introduction}

83: 

84: Scheduling problems have been studied extensively from the point of

85: view of the objectives of the enterprise that stands to gain from the

86: completion of a set of jobs.  We take a new look at the problem from

87: the point of view of the workers who perform the tasks that earn the

88: company its profits.  In fact, it is natural to expect that some

89: employees may lack the motivation to perform at their peak levels of

90: efficiency, either because they have no stake in the company's profits

91: or because they are simply lazy.  

92: 

93: The following example illustrates

94: the situation facing a ``typical'' office worker, who may be one small

95: cog in a large bureaucracy:

96: 

97: \begin{quote}

98: {\em Example.}\quad It is 3:00~p.m., and Dilbert goes home at

99: 5:00~p.m.  Dilbert has two tasks that have been given to him: one

100: requires 10 minutes, the other requires an hour.  If there is a task

101: in his ``in-box,'' Dilbert must work on it, or risk getting fired.

102: However, if he has multiple tasks, Dilbert has the freedom to choose

103: which one to do first.  He also knows that at~3:15, another task will

104: appear --- a 45-minute personnel meeting.  If Dilbert begins the

105: 10-minute task first, he will be free to attend the personnel meeting

106: at~3:15 and then work on the hour-long task from~4:00 until~5:00.  On

107: the other hand, if Dilbert is part way into the hour-long job at~3:15,

108: he may be excused from the meeting.  After finishing the 10-minute job

109: by~4:10, he will have 50~minutes to twiddle his thumbs, iron his tie,

110: or enjoy engaging in other mindless trivia.  Naturally, Dilbert

111: prefers this latter option.

112: \end{quote}

113: 

114: An historical example of a situation where it proved crucial to

115: schedule tasks inefficiently is documented in the book/movie {\em

116:   Schindler's List\/}~\cite{Keneally-82}.  It was essential for the

117: workers and management of Schindler's factory to appear to be busy at

118: all times in order for the factory to stay in operation, but they

119: simultaneously sought to minimize their contribution to the German war

120: effort.

121: 

122: These examples illustrate a general and natural type of scheduling

123: problem, which we term the ``Lazy Bureaucrat Problem'' (LBP); the goal

124: of the LBP is to schedule jobs as {\em inefficiently\/} (in some

125: sense) as possible.  There exists a vast literature on scheduling; see

126: e.g., some of the recent

127: surveys~\cite{KargerSteinWein:1997,LLKS-93,Pinedo-95}.  The LBP

128: studies these traditional problems ``in reverse.''  Several other

129: combinatorial optimization problems have also been studied in reverse,

130: leading, e.g., to maximum

131: TSP~\cite{BarvinokJWW1998,Fekete1999,HassinRubinstein1998,KosarajuParkStein1994},

132: maximum cut~\cite{GoemansWilliamson1995}, and longest

133: path~\cite{KargerMotwaniRamkumar1997}; such inquiries often lead to a

134: better understanding of the structure and algorithmic complexity of

135: the original optimization problem.

136: 

137: \subsection{The Model} 

138: \label{subsec-model}

139: 

140: In this paper we schedule a set of jobs $1 \ldots n$ having processing

141: times (lengths) $t_1 \ldots t_n$ respectively.  Job $i$ {\em

142:   arrives\/} at time $a_i$ and has its {\em deadline\/} at time $d_i$.

143: We assume throughout this paper that $t_i$, $a_i$, and $d_i$ have

144: nonnegative integral values.  The jobs have {\em hard deadlines,\/}

145: meaning that each job $i$ can only be executed during its allowed

146: interval $I_i = [a_i, d_i]$; we also call $I_i$ the job's {\em

147:   window\/}. We let $c_i=d_i-t_i$ denote the {\em critical time\/} of

148: job~$i$; job~$i$ must be started by time $c_i$ if there is any chance

149: of completing it on time.

150: 

151: The jobs are executed on a single processor, the (lazy) {\em

152: bureaucrat\/}.  The bureaucrat executes only one job at a time.  

153: 

154: \subsubsection{Busy Requirement}

155: 

156: The bureaucrat chooses a subset of jobs to execute.  Since the

157: bureaucrat's goal is to minimize his effort, he would prefer to remain

158: idle all the time and to leave all the jobs unexecuted.  However, this

159: scenario is forbidden by what we call the {\em busy requirement,\/}

160: which stipulates that the bureaucrat work on an {\em executable\/}

161: job, if any executable jobs exist.  A job is {\em executable\/} if the

162: constraints of the LBP allow the job to be run.  Thus, in the

163: nonpreemptive setting a job $j $ is executable as long as it begun

164: during the interval $[a_j, c_j]$, at which point it is run to

165: completion.  In the preemptive setting the constraints that govern

166: whether or not a job can be executed are more complicated; see

167: Section~\ref{sec:preempt}.

168: 

169: \subsubsection{Objective Functions}

170: 

171: In traditional scheduling problems, if it is impossible to complete

172: the set of all jobs by their deadlines, one typically optimizes

173: according to some objective, e.g., to maximize a weighted sum of

174: on-time jobs, to minimize the maximum lateness of the jobs, or to

175: minimize the number of late jobs.  For the LBP we consider three

176: different objective functions, which naturally arise from the

177: bureaucrat's goal of inefficiency:

178: %

179: \begin{enumerate}

180: 

181: \item

182: {\em Minimize the total amount of time spent working\/} ---

183: This objective naturally appeals to a ``lazy'' bureaucrat.

184: 

185: \item

186: {\em Minimize the weighted sum of completed jobs\/} ---  In this paper

187: we usually assume that the weight of job $i$ is its length, $t_i$;

188: however, other weights (e.g., unit weights) are also of interest.

189: This objective appeals to a ``spiteful'' bureaucrat whose goal it is

190: to minimize the fees that the company collects on the basis of his

191: labors, assuming that the fee (in proportion to the task length, or a

192: fixed fee per task) is collected only for those tasks that are

193: actually completed.

194: 

195: \item

196: {\em Minimize the {\em makespan}, the maximum completion

197: time of the jobs\/} ---  This objective appeals to an ``impatient''

198: bureaucrat, whose goal it is to go home as early as possible, at the

199: completion of the last job he is able to complete.  He cares about the

200: number of hours spent at the office, not the number of hours spent

201: doing work (productive or otherwise) at the office.

202: 

203: Note that, in contrast with standard scheduling problems on one

204: processor, the makespan in the LBP varies; 

205: it is a function of which jobs have passed their

206: deadlines and can no longer be executed.

207: 

208: \end{enumerate}

209: 

210: \subsubsection{Additional Parameters of the Model}

211: As with most scheduling problems, additional parameters of the model

212: must be set.  For example, we must explicitly allow or forbid {\em

213: preemption\/} of jobs.  When a job is {\em preempted\/}, 

214: it is interrupted and may be resumed later at no additional cost.  

215: If we forbid preemption, then once a job is begun, 

216: it must be completed without interruptions.

217: 

218: We must also specify whether scheduling occurs {\em on-line\/} or

219: {\em off-line.\/} A scheduling algorithm is 

220: off-line

221: if all the jobs are known to the scheduler at the outset; it is

222: on-line if the jobs are known to the scheduler only as they arrive.

223: In this paper we restrict ourselves to off-line scheduling; we leave

224: the on-line case as an 

225: open problem.

226: 

227: 

228: \begin{figure*}[bt] 

229: \def\RESULT#1{\vbox{\vskip 2pt\hbox{#1}\vskip 0.5pt}}

230: \begin{center}

231: \begin{tabular}{|l||c||c|}

232: \hline

233: {\bf Instance} & {\bf Metric} & {\bf Complexity} \\

234: \hline \hline

235: $\bullet$ Unit-length jobs: $t_1 =  \cdots = t_n = 1$  & \RESULT{1}

236: & \RESULT{Polynomial time} \\

237: \hline

238: $\bullet$ Short job intervals:  \mbox{$\forall i \, \, d_i - a_i < 2 t_i $} & 

239: \RESULT{1-3}  &

240: \RESULT{Pseudo-polynomial time} \\

241: \hline

242: $\bullet$ Ratios $R = O(1)$  and $\Delta = O(1)$ & 

243: \RESULT{1-3}  &

244: \RESULT{Pseudo-polynomial time} \\

245: \hline

246: $\bullet$ Same arrival times: $a_1 = \cdots = a_n$ & \RESULT{1-3}  

247: &

248: \RESULT{Pseudo-polynomial time} \\

249: \RESULT{~~~~}  & \RESULT{~~~~}  & 

250:  \RESULT{(Weakly) NP-complete} \\

251: \RESULT{~~~~} & 

252: \RESULT{~~~~~~} & 

253: \RESULT{Hard to approximate} \\

254: \hline

255: $\bullet$ Job size ratio $\Delta = O(1)$

256: & \RESULT{1-3}

257: &

258:  \RESULT{Strongly NP-complete} \\

259: \RESULT{~~~~}  & \RESULT{~~~~}  &

260:  \RESULT{Hard to approximate} \\

261: \hline

262: $\bullet$ General LBP

263: & \RESULT{1-3}

264: &

265:  \RESULT{Strongly NP-complete} \\

266: \RESULT{~~~~}  & \RESULT{~~~~}  &

267:  \RESULT{Hard to approximate} \\

268: \hline

269: \end{tabular}

270: \end{center}

271: \caption{

272:   Summary of our nonpreemptive results; see Section~\ref{nonpreempt}.}

273: \label{tab:nonpreemtive-results}

274: \end{figure*}

275: 

276: \subsection{Our Results}

277: 

278: In this paper, we introduce the Lazy Bureaucrat Problem and develop

279: algorithms and hardness results for several versions of the LBP.  {From}

280: these results, we derive some general characteristics of this new

281: class of scheduling problems and describe (1) situations in which

282: traditional scheduling algorithms extend to the LBP and (2)

283: situations in which these algorithms no longer apply.  

284: 

285: 

286: \subsubsection{No Preemption}

287: We prove that the LBP is NP-complete, as is often the case for

288: traditional scheduling problems. Thus, we focus on special cases to

289: study exact algorithms.  When all

290: jobs have unit size, optimal schedules can be found in polynomial

291: time.

292: The following three cases have pseudo-polynomial algorithms:

293: (1) when each job $i$'s interval $I_i$ is less 

294: than twice the processing time of job $i$;

295: (2) when the ratios of interval length to job length and longest job

296: to shortest job are both bounded; and (3) when all jobs arrive in the

297: system at the same time.

298: These last scheduling problems are solved using dynamic programming

299: both for Lazy Bureaucrat and traditional metrics.  Thus, in these

300: settings, the Lazy Bureaucrat metrics and traditional metrics are

301: solved using similar techniques.

302: 

303: {From} the point of view of approximation, however, the standard and

304: Lazy Bureaucrat metrics behave differently.  Standard metrics

305: typically allow polynomial-time algorithms having good approximation

306: ratios, whereas we show that the Lazy Bureaucrat metrics are difficult

307: to approximate.  This hardness derives more from the busy requirement

308: and less from the particular metric in question, that is the busy

309: requirement appears to render the problem substantially more

310: difficult.  (Ironically, even in standard optimization problems, the

311: management often imposes this requirement because it intuitively

312: appears desirable.)

313: 

314: \begin{figure*}[bt] %Htp

315: \def\RESULT#1{\vbox{\vskip 2pt\hbox{#1}\vskip 0.5pt}}

316: \begin{center}

317: \begin{tabular}{|c|c||l|}

318: \hline

319: {\bf Preemption} & {\bf Metrics} & {\bf Complexity}\\

320: \hline \hline

321: \RESULT{I} & \RESULT{1-3} &

322: \RESULT{Polynomial Time} \\

323: \hline

324: \RESULT{II} & \RESULT{1-3} &

325: \RESULT{(Weakly) NP-complete,  even when $a_0 = \cdots = a_n$} \\

326: \hline

327: \RESULT{III} & \RESULT{1-3} &

328: \RESULT{(Weakly) NP-complete,  even when \mbox{$a_0 = \cdots = a_n$}} \\

329: \RESULT{~~~} & \RESULT{~~~} & \RESULT{Hard to approximate} \\

330: \hline

331: \end{tabular}

332: \end{center}

333: \caption{Summary of our preemptive results; see Section~\ref{sec:preempt}.}

334: \label{tab:preemtive-results}

335: \end{figure*}

336: 

337: 

338: \subsubsection{Preemption}

339: 

340: The busy requirement dictates that the worker must stay busy while

341: work is in the system. If the model allows preemption we must specify

342: under what conditions a job can be interrupted or resumed. We

343: distinguish three versions of the preemption rules, which we list from

344: most permissive to most restrictive. In particular,

345: workers are constrained to execute the following jobs:

346: \begin{description}

347: \item[] (I) any job that has arrived and is before its deadline,

348: \item[] (II) any job that has arrived and for which there is still

349:   time to complete it before its deadline, or

350: \item[] (III) any job that has arrived, but with the constraint that

351:   if it is started, it must eventually be completed.

352: \end{description}

353: 

354: We consider all three metrics and all three versions of preemption. We

355: show that, for all three metrics, version~I is polynomially solvable,

356: and version~III is NP-complete. Many of the hardness results for no

357: preemption carry over to version~III. 

358: 

359: Our main results are for version~II. We show that the general problem

360: is NP-complete. Then, we focus on minimizing the makespan in two

361: complementary special cases: 

362: %

363: \begin{enumerate}

364: 

365: \item

366: All jobs have a common arrival time and arbitrary deadlines.

367: 

368: \item

369:  All jobs have a common deadline and arbitrary arrival times.  

370: \end{enumerate}

371: 

372: \noindent

373: We show that the first problem is

374: NP-complete, whereas the second problem can be solved in polynomial

375: time.

376: 

377: These last results illustrate a curious feature of the LBP. One can

378: convert one special case into the other by reversing the direction of

379: time. In the LBP, unlike many scheduling settings, this reversing of

380: time changes the complexity of the problem.

381: 

382: \begin{figure*}[bt] 

383: \def\RESULT#1{\vbox{\vskip 2pt\hbox{#1}\vskip 0.5pt}}

384: \begin{center}

385: \begin{tabular}{|l||c|c||c|}

386: \hline

387: {\bf Instance} & 

388: {\bf Preemption} & 

389: {\bf Metric} & 

390: {\bf Complexity}\\

391: \hline \hline

392: $\bullet $ Arbitrary deadlines  &

393: \RESULT{II} &

394: \RESULT{3} & 

395: \RESULT{(Weakly) NP-Complete} \\

396: %

397: $\bullet $ Identical job arrivals: &

398: \RESULT{~~~} & 

399: \RESULT{~~~} & 

400: \RESULT{Hard to approximate} \\

401: %

402: ~~~~$a_1 =  \cdots = a_n = 0$ & 

403: \RESULT{~~~} & 

404: \RESULT{~~~} & 

405: \RESULT{~~~} \\

406: \hline 

407: %

408: $\bullet $ Arbitrary job arrivals &

409: \RESULT{II} &

410: \RESULT{3} & \RESULT{Polynomial time} \\

411: $\bullet $ Identical deadlines: & 

412: \RESULT{~~~} &

413: \RESULT{~~~} &

414: \RESULT{~~~} \\

415: %

416: ~~~~$d_1 =  \cdots = d_n = D$  &

417: \RESULT{~~~} &

418: \RESULT{~~~} &

419: \RESULT{~~~} \\

420: \hline

421: \end{tabular}

422: \end{center}

423: \caption{

424:   Illustration that reversal of time changes complexity; see

425:   Section~\protect\ref{sec:pmtn-makespan}.}

426: \label{tab:time-reversal}

427: \end{figure*}

428: 

429: \subsection{Related Work on the LBP}

430: 

431: Recently Hepner and Stein~\cite{HepnerStein2002}

432: published a pseudo-polynomial-time algorithm

433: for minimizing the makespan subject to preemption constraint II,

434: thus resolving an open problem from

435: an earlier version of this paper~\cite{ArkinBenderMitchellSkiena99}.

436: They also extend the LBP to the parallel setting, 

437: in which there are multiple bureaucrats.

438: 

439: 

440: \section{LBP: No Preemption}

441: \label{nonpreempt}

442: 

443: 

444: In this section, we assume that no job can be preempted: if a job is

445: started, then it is performed without interruption until it completes.

446: We show that the Lazy Bureaucrat Problem (LBP) without preemption is

447: strongly NP-complete and is not approximable to within any factor for

448: the three metrics we consider.  These hardness results distinguish our

449: problem from traditional scheduling metrics, which can be approximated

450: in polynomial time, as proved in~\cite{BGNS-99}.  We show, however,

451: that several special cases of the problem have pseudo-polynomial-time

452: algorithms, using applications of dynamic programming.

453: 

454: \subsection{Hardness Results}

455: \label{sec:hard}

456: 

457: We begin by describing the relationship

458: between the three different objective functions 

459: from Section~\ref{sec:hard}

460: in the case of no preemption.  

461: 

462: The problem of minimizing the

463: total work (objective function 1) is a special case of the problem of 

464: minimizing the weighted sum of completed jobs 

465: (objective function 2), because without preemption 

466: every job that is

467: executed must be completed.  (The weights become the job lengths.)

468: Furthermore, if all jobs have

469: the same arrival time, say time zero, then the 

470: two objectives 

471: minimizing the total amount of time spent working

472: and minimizing  the makespan (go home early) are equivalent

473: (objective functions 1 and 3), since

474: no feasible schedule will have any gaps.

475: Our first hardness theorem

476: applies therefore to all three objective functions

477: from Section~\ref{sec:hard}.

478: 

479: \begin{thm}

480: \label{thm:hard1}

481: The Lazy Bureaucrat Problem with no preemption is (weakly) NP-complete

482: for objective functions (1)-(3), 

483: and is not approximable to within any fixed factor, even when all arrival

484: times are the same.

485: \end{thm}

486: 

487: \begin{pf}

488: We use a reduction from the {\sc Subset Sum} problem~\cite{GJ-79}:

489: Given a set of

490: integers $S=\{x_1,x_2,\ldots,x_n\}$ and a target integer $T$, does there 

491: exist a subset $S'\subseteq S$, such that $\sum_{x_i\in S'}x_i=T$?

492: 

493: We construct an instance of the LBP having $n+1$ jobs, each having

494: release time zero ($a_i=0$ for all $i$). For $i=1\ldots,n$, job $i$

495: has processing time $t_i=x_i$ and deadline $d_i=T$.  Job $n+1$ has

496: processing time $t_{n+1}=1+\sum_{x_i\in S}x_i$ and deadline

497: $d_{n+1}=T+t_{n+1}-1$; thus, job $n+1$ can be started at time $T-1$ or

498: earlier.  Because job $n+1$ is so long, the bureaucrat wants to avoid executing

499: it, but can do so if and

500: only if he selects a subset of jobs from $\{1,\ldots,n\}$ to execute

501: whose lengths sum to exactly~$T$.

502: In summary, the large job $n+1$ is executed if and only if

503: the subset problem is solved exactly and executing the long job

504: leads to a schedule whose makespan (i.e., total work executed)

505: is not within any fixed factor of the optimal solution.

506: \end{pf}

507: 

508: 

509: We now show that the LBP with no preemption

510: is strongly NP-complete.  

511: As we will show in Section~\ref{pseudo}, the LBP from Theorem~\ref{thm:hard1}

512: when all arrival times are equal,

513: has a pseudo-polynomial-time algorithm.

514: However, if arrival times and deadlines are arbitrary integers, the

515: problem becomes strongly NP-complete. 

516: Thus, the following theorem

517: subsumes Theorem~\ref{thm:hard1} when arrival times and deadlines

518: our unconstrained, whereas

519: Theorem~\ref{thm:hard1} is more generally applicable.

520: 

521: \begin{thm}

522: \label{thm:hard2}

523: The Lazy Bureaucrat Problem with no preemption is strongly

524: NP-complete for objective functions (1)--(3), and is not approximable to within any fixed factor.

525: \end{thm}

526: 

527: \begin{pf}

528: Clearly the problem is in NP, since any solution can be represented

529: by an ordered list of jobs, given their arrival times.

530: To show hardness,

531: we use a reduction from the {\sc 3-Partition} problem~\cite{GJ-79}:

532: Given a

533: set $S=\{x_1,\ldots,x_{3m}\}$ of $3m$ positive integers and a positive

534: integer bound $B$ such that $B/4<x_i<B/2$, for $i=1,\ldots,3m$ and

535: $\sum_i x_i=mB$, does there exist a partitioning of $S$ into $m$

536: disjoint sets, $S_1,\ldots,S_m$, such that for $i=1,...,m$,

537: $\sum_{x_j\in S_i} x_j=B$? (Note that, by the assumption that

538: $B/4<x_i<B/2$, each set $S_i$ must contain exactly 3 elements.)

539: 

540: Objective function (1)

541: is a special case of

542: objective function (2)

543: because without preemption, any job that is begun must be completed.

544: Furthermore, hard instances will be designed so that there are no gaps,

545: ensuring that the optimal solution for objective function (1)

546: is also the optimal solution for objective function (3).

547: 

548: 

549: We construct an instance of the LBP containing three classes of jobs:

550: 

551: \begin{itemize}

552: \item {\em Element jobs\/} --- We define one ``element job'' corresponding to

553: each element $x_i\in S$, having arrival time~$0$, deadline

554: $d_i=(m-1)+mB$, and processing time~$x_i$.

555: 

556: \item {\em Unit jobs\/} --- 

557:  We define $m-1$ ``unit'' jobs, each of length

558: $1$. The $i$-th unit job (for $i=1,\ldots,m-1$) has arrival time

559: $i(B+1)-1$ and deadline $i(B+1)$. Note that for these unit-length jobs

560: we have $d_j-a_j=1$; thus, these jobs must be processed immediately

561: upon their arrival, or not at all.

562: 

563: \item {\em Large job\/} --- 

564: We define one ``large'' job of length $L>(m-1)+mB$,

565: arrival time $0$, and deadline $L+(m-2)+mB$. Note that in order to

566: complete this job, it must be started at time $(m-2)+mB$ or before.

567: \end{itemize}

568: 

569: As in the proof of Theorem~\ref{thm:hard1}, the lazy bureaucrat wants

570: to avoid executing the long job, but can do so if and only if all

571: other jobs are actually executed.  Otherwise, there will be a time

572: when the large job is the only job in the system and the lazy

573: bureaucrat will be forced to execute it.  Thus, the unit jobs must be

574: done immediately upon their arrival, and the element jobs must fit in

575: the intervals between the unit jobs.  Each such interval between

576: consecutive unit jobs is of length exactly $B$.  Refer to

577: Figure~\ref{fig:no-preempt-hard}. In summary, the long job is not

578: processed if and only if all of the element and unit jobs can be

579: processed before their deadlines, which happens if and only if the

580: corresponding instance of {\sc 3-Partition} is a ``yes'' instance.

581: Note that since $L$ can be as large as we want, this also implies that

582: no polynomial-time approximation algorithm with any fixed

583: approximation bound can exist, unless P=NP.  

584: \end{pf}

585: 

586: \begin{figure}[htbp]

587: \begin{center}

588: \input{fig1.pstex_t}

589: \end{center}

590: \caption{

591:   Proof of hardness of LBP with no preemption and arbitrary arrival

592:   times.}

593: \label{fig:no-preempt-hard}

594: \end{figure}

595: 

596: 

597: \subsection{Algorithms for Special Cases}

598: \label{pseudo}

599: 

600: \subsubsection{Unit-Length Jobs}

601: 

602: Consider the special case of the LBP in which all

603: jobs have unit processing times.

604: (Recall that all inputs are assumed to be integral.)

605: The Latest Due Date (LDD) scheduling policy 

606: selects the job in the system having the latest deadline.

607: Note that this policy in nonpreemptive for unit-length jobs, since

608: all jobs have integral arrival times.

609: 

610: 

611: \begin{thm}

612: Consider the Latest Deadline First scheduling policy 

613: when jobs have unit lengths and all inputs are integral.

614: The LDD scheduling policy

615: minimizes the amount of executed work.

616: \end{thm}

617: 

618: \begin{pf}

619: Assume by contradiction that no optimal schedule is LDD.  We use an

620: exchange argument. 

621: Consider an optimal (non-LDD) schedule that has

622: the fewest pairs of jobs executed in non-LDD order.  

623: The schedule must have two neighboring jobs $i,j$ such that $i<j$ in the

624: schedule but $D_i<D_j$, and $j$ is in the system when

625: $i$ starts its execution. Consider the first such pair of jobs. 

626: There are two cases:

627: 

628: (1) The new schedule with $i$ and $j$ switched, is feasible. It

629: executes no more work than the optimal schedules, and is therefore

630: also optimal.

631: 

632: (2) The schedule with $i$ and $j$ switched is not feasible. This

633: happens if $i$'s deadline has passed.  If no job is in the system

634: to replace $i$, then we obtain a better schedule than the optimal schedule

635: and reach a contradiction.

636: Otherwise, we replace $i$ with the other job and repeat the switching

637: process.

638: 

639: We obtain a schedule executing no more work than an optimal

640: schedule, but with fewer pairs of jobs in non-LDD order, a contradiction.

641: \end{pf}

642: 

643: 

644: \subsubsection{Narrow Windows}

645: Consider now the version in which jobs are large in comparison with

646: their intervals,  that is, the intervals are ``narrow.''

647: Let $R$ be a bound on the ratio of window length to job length; i.e.,

648: for each job $i$, $d_i-a_i< R \cdot t_i$. 

649: We show that a pseudo-polynomial algorithm exists for the case of

650: sufficiently narrow windows, that is,  when $R \leq 2$.

651: 

652: \begin{lem}

653: \label{lem:unique-ordering}

654: Assume that for each job $i$, $d_i-a_i<2t_i$.

655: Then, if job $i$ can be

656: scheduled before job $j$, then job $j$ cannot be scheduled before

657: job~$i$.

658: \end{lem}

659: 

660: \begin{pf}

661: We rewrite the assumption: for each $i$, $d_i-t_i< t_i+a_i$.

662: The fact that job $i$ can be scheduled before job $j$ is equivalent to

663: the statement that $a_i+t_i\leq d_j-t_j$, since the earliest that job

664: $i$ can be completed is at time $a_i+t_i$ and the latest that job $j$

665: can be started is at time $d_j-t_j$.  Combining these inequalities,

666: we obtain

667: %

668: $$a_j+t_j> d_j-t_j\geq a_i+t_i > d_i-t_i,$$

669: %

670: which implies that job $j$ cannot be scheduled before job~$i$.

671: \end{pf}

672: 

673: \begin{cor}

674: Under the assumption that $d_i-a_i<2t_i$ for each $i$, the ordering

675: of any subset of jobs in a schedule is uniquely determined.

676: \end{cor}

677: 

678: \begin{thm}

679: Suppose that for each job $i$, $d_i-a_i< 2 t_i$.

680: Let $K=\max_i d_i$.

681: Consider the problem of minimizing 

682: objective functions~(1)-(3) 

683: from Section~\ref{subsec-model}, 

684: in the nonpreemptive setting.

685: Then the LBP can be solved in $O(n K \max(n,K))$ time.

686: \label{fixed-window-DP}

687: \end{thm}

688: 

689: \begin{pf}

690: We use dynamic programming to find the shortest path in a 

691: directed acyclic graph (DAG).

692: There are $O(nK^2)$ states the system can enter.

693: Let $(i,j,\tau)$ denote the state of the system when the processor

694: begins executing the $j$-th unit of work of job $i$ at time $\tau$.

695: Thus, $i=1,...,n$, $j=1,...,t_i$, and $\tau =0,...,K$. Transitions

696: from state to state are defined according to the following rules:

697: 

698: \begin{enumerate}

699: \item No preemption: once a job is begun, it must be completed without

700:   interruptions.

701: \item When a job is completed at time $\tau$, another job must begin

702:   immediately if one exists in the system. (By

703:   Lemma~\ref{lem:unique-ordering}, we know this job has not yet been

704:   executed.) Otherwise, the system is idle and begins executing a job

705:   as soon as one arrives.

706: \item State $(i,t_i,\tau)$ is an end state if and only if when job $i$

707:   completes at time $\tau$, no jobs can be executed subsequently.

708: \item The start state has transitions to the jobs $(i,1,0)$ that

709:   arrive first.

710: \end{enumerate}

711: 

712: The goal of the dynamic program is to find the length of a shortest

713: path from the start state to an end state.  Depending on how we assign

714: weights to the edges we can force our algorithm to minimize all three

715: metrics from Section~\ref{subsec-model}.  To complete the time

716: analysis, note that only $n K$ of the $nK^2$ states have more than

717: constant outdegree, and these states each have outdegree bounded

718: by~$n$.

719: \end{pf}

720: 

721: For $R>2$ we know of no efficient algorithm without additional

722: conditions.

723: Let $W$ be a bound on the ratio of longest window to shortest window, and let

724: $\Delta$ be a bound on the ratio of the longest job to the shortest job.

725: Note that bounds on $R$ and $\Delta$ imply a bound on $W$, and

726: bounds on $R$ and $W$ imply a bound on $\Delta$.

727: However, a bound on $\Delta$ alone is not sufficient for a 

728: pseudo-polynomial-time algorithm.

729: 

730: \begin{thm}

731: \label{thm:bounded-Delta}

732: Even with a bound on the ratio $\Delta$, the LBP with no preemption

733: is strongly NP-complete for objective functions~(1)-(3).

734: It cannot be approximated to within  a factor of $\Delta - \epsilon$, 

735: for any $\epsilon>0$, unless P=NP.

736: \end{thm}

737: 

738: \begin{pf}

739: Modify the reduction from 3-partition of Theorem~\ref{thm:hard2},

740: by changing all the fixed ``unit'' jobs to have length $B/3$,

741: and adjust the arrival times and deadlines accordingly.

742: 

743: Instead of one very long job

744: as in the proof from Theorem~\ref{thm:hard2},

745: we create a sequence of bounded-length jobs that serve the same purpose.

746: One unit before the deadline of the ``element'' jobs

747: (see Theorem~\ref{thm:hard2}) a sequence of longer jobs

748: $\ell_1,\ldots,\ell_m$ arrives.  Each job $\ell_i$ entirely fills its window

749: and so can only be executed directly when it arrives.  Job  $\ell_{i+1}$

750: arrives at the deadline of job  $\ell_i$.  In addition, a sequence of

751: shorter jobs $s_1,\ldots,s_m$ arrives, where each

752: shorter job $s_i$ also entirely fills its window and can only be executed 

753: when it arrives.  Shorter job $s_i$ overlaps $\ell_i$ and $\ell_{i+1}$; it

754: arrives one unit before the deadline of

755: job $\ell_i$. 

756: Jobs $s_i$ have length $B/4$ and jobs $\ell_i$ have length $\Delta\cdot B/4$.

757:  Thus, if all the jobs comprising the

758: $3$-partition problem can be executed, jobs $\ell_1,\ldots,\ell_m$

759: will be avoided by executing jobs $s_1,\ldots,s_m$.  Otherwise, jobs

760: $\ell_1,\ldots,\ell_m$ must be executed.  The index of $m$ can be adjusted

761: to any $\epsilon$.

762: \end{pf}

763: 

764: Bounds on both $\Delta$ and $R$ are sufficient to yield a

765: pseudo-polynomial algorithm:

766: 

767: \begin{thm}

768: Let $K = \max_i d_i$.

769: Given bounds on $R$ and $\Delta$,

770: the Lazy Bureaucrat Problem with no preemption

771: can be solved in $O(K\cdot n^{4 R\lg \Delta})$

772: for objective functions~(1)-(3).

773: \end{thm}

774: 

775: \begin{pf}

776: We modify the

777: dynamic programming algorithm of Theorem~\ref{fixed-window-DP}

778: for this more complex situation.

779: The set of jobs potentially available to work on in a given schedule

780: at time $\tau$ are the jobs $j$ that have not yet been executed, for which 

781: $d_j - t_j \geq \tau$.

782: Our state space will encode the {\em complement\/} of this set for each time 

783: $\tau$,

784: specifically, the set of jobs that were executed earlier but

785: could otherwise have been executed at time $\tau$.

786: 

787: The bounds on $R$ and $\Delta$ together imply an upper bound on 

788: the number of subsets of jobs active at time 

789: $\tau$ that could have been executed

790: prior to time $\tau$.

791: Let $d_{\min}$ be the length of the shortest job potentially 

792: active at time $\tau$.

793: We can partition all potentially active jobs into $\lg \Delta$ classes,

794: where the $j$th class consists of the jobs of size

795: greater than or equal to $2^{j-1} d_{\min}$

796: and less than  $2^j d_{\min}$.

797: The earliest possible arrival time of any class-$j$ job

798: is $\tau - 2^j d_{\min} R$, since each job has an $R$-bounded window.

799: Only $2^j d_{\min} R / 2^{j-1} d_{\min} = 2R$ jobs from class $j$

800: can be executed within this window.

801: Summing over all the classes implies that at most $2 R \lg \Delta$ jobs

802: potentially active at time $\tau$ could have been executed in a non-preemptive

803: schedule by time $\tau$.

804: As before the choice of weights on the edges

805: determines the metric that is optimized.

806: 

807: The time bound on the running time 

808: follows by observing that each of the $K n^{(2 R \lg \Delta)}$

809: states has outdegree at most $2 R \lg \Delta$.

810: \end{pf}

811: 

812: \subsubsection{Jobs Having a Common Release Time}

813: 

814: In the next version of the problem all jobs are released at time zero,

815: i.e., $a_i=0$ for all $i$.  This problem can be solved in

816: pseudo-polynomial time by dynamic programming, specifically, reducing

817: the problem to that of finding the shortest path in a directed acyclic

818: graph.  The dynamic programming works because of the following

819: structural result: There exists an optimal schedule that executes the

820: jobs Earliest Due Date (EDD).

821: 

822: In fact this problem is a special case of the following general

823: problem: Minimizing the weighted sum of jobs not completed by their

824: deadlines.  A similar problem was solved by~\cite{LawlerMoore69}, using the

825: same structural result.

826: 

827: 

828: \begin{thm}

829: The LBP can be solved in pseudo-polynomial time 

830: for all three metrics when all jobs have

831: a common release time.

832: Specifically, let $K = \max_i d_i$; then the running time is

833: $O(K\cdot n).$

834: \end{thm}

835: 

836: \section{LBP: Preemption}

837: \label{sec:preempt}

838: 

839: In this section we consider the Lazy Bureaucrat Problem in which 

840: jobs may be preempted: a job in progress can be set aside, while another

841: job is processed, and then possibly resumed later.  It is important to

842: distinguish among different constraints that specify which jobs

843: are available to be processed.  We consider three natural choices of such 

844: constraints:

845: 

846: \begin{description}

847: \item[Constraint I:] In order to work on job $i$ at time $\tau$, we

848: require only that the current time $\tau$ lies within the job's

849: interval $I_i$: $a_i\leq \tau\leq d_i$.

850: 

851: \item[Constraint II:] In order to work on job $i$ at time $\tau$, we

852: require not only that the current time $\tau$ lies within the job's

853: interval $I_i$, but also that the job has a {\em chance\/} to be

854: completed, e.g., if it is processed without interruption until

855: completion.

856: 

857: This condition is equivalent to requiring that $\tau\leq c'_i$, where

858: $c'_i=d_i-t_i+y_i$ is the {\em adjusted critical time\/} of job $i$:

859: $c'_i$ is the latest possible time to start job $i$, in order to meet

860: its deadline $d_i$, given that an amount $y_i$ of the job has already

861: been completed.

862: 

863: 

864: 

865: \item[Constraint III:] In order to work on job $i$, we require that

866: $\tau\in I_i$.  Further, we require that any job that is started

867: is eventually completed.

868: 

869: \end{description}

870: 

871: 

872: 

873: We divide this section into subsections,

874: where each subsection considers one of 

875: the three objective functions (1)--(3)

876: from Section~\ref{subsec-model}, in

877: which the goals are to minimize (1) the total time working (regardless of

878: which jobs are completed), (2) the weighted sum of completed jobs, or

879: (3) the makespan of the schedule (the ``go home'' time).

880: For each metric we see that the constraints on preemption

881: can dramatically affect the complexity of the problem.

882: 

883: Constraint III makes the LBP with preemption quite similar

884: to the LBP with no preemption. In fact, if all jobs arrive at the same

885: time ($a_i=0$ for all $i$), then the three objective functions are

886: equivalent, and the problem is hard:

887: 

888: \begin{thm}

889: \label{hard-pre-3}

890: The LBP with preemption, under constraint~III (one must complete any

891: job that is begun), is (weakly)

892: NP-complete and hard to approximate for all three objective

893: functions.

894: \end{thm}

895: 

896: \begin{pf}

897: We use the same reduction as the one given in the proof of

898: Theorem~\ref{thm:hard1}.  Note that any schedule for an instance given

899: by the reduction, in which all jobs processed must be completed

900: eventually, can be transformed into an equivalent schedule with no

901: preemptions. This makes the problem of finding an optimal schedule

902: with no preemption equivalent to the problem of finding an optimal

903: schedule in the preemptive case under constraint III.  

904: \end{pf}

905: 

906: Note that we

907: cannot use a proof similar to that of Theorem~\ref{thm:hard2} to show

908: that this problem is strongly NP-complete, since preemption can lead

909: to improved schedules in that instance.

910: 

911: \subsection{Minimizing Total Time Working}

912: 

913: \begin{thm}

914: \label{thm:preempt-I.1}

915: The LBP with preemption, under constraint~I (one can work on any job

916: in its interval) and objective~(1) (minimize total time

917: working), is polynomially solvable.

918: \end{thm}

919: 

920: \begin{pf}

921: The algorithm schedules jobs according to latest due date

922: (LDD), in which at all times the job in the system with the latest

923: deadline is being processed, with ties broken arbitrarily.  An

924: exchange argument shows that this is optimal.  Suppose there is an

925: optimal schedule that is not LDD. Consider the first time in which an

926: optimal schedule differs from LDD, and let OPT be an optimal schedule

927: in which this time is as late as possible. Let OPT be executing a

928: piece of job $i$, $p_i$ and LDD executes a piece of job $j$, $p_j$. We

929: know that $d_i<d_j$. We want to show that we can replace the first

930: unit of $p_i$ by one unit of $p_j$, contradicting the choice of OPT,

931: and thereby proving the claim. If in OPT, job $j$ is not completely

932: processed, then this swap is feasible, and we are done. On the other

933: hand, if all of $j$ is processed in OPT,  such a swap causes a unit of

934: job $j$ later on to be removed, leaving a gap of one unit. If this gap

935: cannot be filled by any other job piece, we get a schedule with less

936: work than OPT, which is a contradiction. Therefore assume the gap can

937: be filled, possibly causing a later unit gap. Continue this process,

938: and at its conclusion, either a unit gap remains contradicting the

939: optimality of OPT, or no gaps remain, contradicting the choice of OPT.

940: \end{pf}

941: 

942: 

943: 

944: \begin{thm}

945: The LBP with preemption, under constraint~II (one can only work on

946: jobs that can be completed) and objective~(1) (minimize total time

947: working), is (weakly) NP-complete.

948: \end{thm}

949: 

950: \begin{pf}

951: If all arrival times are the same, then this problem is equivalent to

952: the one in which the objective function is to minimize the makespan,

953: which is shown to be NP-complete in Theorem~\ref{thm:preempt-II.3}. 

954: \end{pf}

955: 

956: 

957: \subsection{Minimizing Weighted Sum of Completed Jobs}

958: 

959: \begin{thm}

960: The LBP with preemption, under constraint~I (one can work on any job

961: in its interval) and objective~(2) (minimize the weighted sum of

962: completed jobs), is polynomially solvable.

963: \end{thm}

964: 

965: \begin{pf}

966: Without loss of generality, assume that jobs $1,\ldots,n$ are

967: indexed in order of increasing deadlines. We show how to decompose the

968: jobs into separate components that can be treated independently.

969: Schedule the jobs according to EDD (if a job is executing and its

970: deadline passes, preempt and execute the next job).  Whenever there is

971: a gap (potentially of size zero), where no jobs are in the system, the

972: jobs are divided into separate components that can be scheduled

973: independently and their weights summed.  Now we focus on one such set

974: of jobs (having no gaps).  We modify the EDD schedule by preempting a

975: job $\epsilon$ units of time before it completes.  Then we move the

976: rest of the jobs of the schedule forward by $\epsilon$ time units and

977: continue to process.  At the end of the schedule, there are two

978: possibilities.  (1) the last job is interrupted because its deadline

979: passes; in this case we obtain a schedule in which no jobs are

980: completed; (2) the last job completes and in addition

981: all other jobs whose deadlines have not passed are also forced to

982: complete. 

983: 

984: The proof is completed by noting the following:

985: %

986: \begin{itemize}

987: \item There is an optimal schedule that completes all of its jobs at

988:   the end; and

989: \item The above schedule executes the maximum amount of work possible.

990:   (In other words, EDD (``minus $\epsilon$'') allows one to execute

991:   the maximum amount of work on jobs 1 through $i$ without completing

992:   any of them.)

993: \end{itemize}

994: \end{pf}

995: 

996: 

997: \begin{thm}

998: The LBP with preemption, under constraint~II (one can only work on

999: jobs that can be completed) and objective~(2) (minimize the weighted

1000: sum of completed jobs), is (weakly) NP-complete.

1001: \end{thm}

1002: 

1003: \begin{pf}

1004:   Consider the LBP under constraint~II, where the objective is to

1005:   minimize the makespan.  The proof of Theorem~\ref{thm:preempt-II.3}

1006:   will have hard instances where all jobs have the same arrival time,

1007:   and where the optimal solution completes any job that it begins.

1008:   Thus, for these instances the metric of minimizing the makespan is

1009:   equivalent to the metric of minimizing the weighted sum of completed

1010:   jobs, for weights proportional to the processing times.

1011: \end{pf}

1012: 

1013: 

1014: \subsection{Minimizing Makespan: Going Home Early}

1015: \label{sec:pmtn-makespan}

1016: 

1017: We assume now that the bureaucrat's goal is to go home as soon as

1018: possible.

1019: 

1020: We begin by noting that if the arrival times are all the same

1021: ($a_i=0$, for all $i$), then the objective (3) (go home as soon as

1022: possible) is in fact equivalent to the objective (1) (minimize total

1023: time working), since, under any of the three constraints I--III, the

1024: bureaucrat will be busy nonstop until he can go home.

1025: 

1026: Observe that if the {\em deadlines\/} are all the same

1027: ($d_i=D$, for all $i$), then the objectives (1) and (3) are quite

1028: different.  Consider the following example.  Job~1 arrives at time

1029: $a_1=0$ and is of length $t_1=2$, job~2 arrives at time $a_2=0$ and is

1030: of length $t_2=9$, job~3 arrives at time $a_3=8$ and is of length

1031: $t_3=2$, and all jobs have deadline $d_1=d_2=d_3=10$.  Then, in order

1032: to minimize total time working, the bureaucrat will do jobs 1 and 3, a

1033: total of 4 units of work, and will go home at time 10.  However, in

1034: order to go home as soon as possible, the bureaucrat will do job~2,

1035: performing 9 units of work, and go home at time 9 (since there is not

1036: enough time to do either job~1 or job~3).

1037: 

1038: \begin{thm}

1039: \label{thm:preempt-I.3}

1040: The LBP with preemption, under constraint~I (one can do any job in its

1041: interval) and objective~(3) (go home as early as possible), is

1042: polynomially solvable.

1043: \end{thm}

1044: 

1045: \begin{pf}

1046: The algorithm is to schedule by Latest Due Date (LDD). 

1047: The proof is similar to the one given in Theorem~\ref{thm:preempt-I.1}.

1048: \end{pf}

1049: 

1050: If instead of constraint~I we impose constraint~II, the problem becomes hard:

1051: 

1052: \begin{thm}

1053: \label{thm:preempt-II.3}

1054: The LBP with preemption, under constraint~II (one can only work on jobs that

1055: can be completed) and objective~(3) (go home as early as possible), is

1056: (weakly) NP-complete, even if all arrival times are the same.

1057: \end{thm}

1058: 

1059: \begin{pf}

1060:  We give a reduction from {\sc Subset Sum}.  Consider an

1061: instance of {\sc Subset Sum} given by a set $S$ of $n$ positive integers,

1062: $x_1$, $x_2,\ldots,x_n$, and target sum $T$.  We construct an instance

1063: of the required version of the LBP as

1064: follows.  For each integer $x_i$, we have a job $i$ that arrives at

1065: time $a_i=0$, has length $t_i=x_i$, and is due at time

1066: $d_i=T+x_i-\epsilon$, where $\epsilon$ is a small constant (it

1067: suffices to use $\epsilon={1\over 3n}$).  In addition, we have a

1068: ``long'' job $n+1$, with length $t_{n+1}>T$, that arrives at

1069: time $a_{n+1}=0$ and is due at time $d_{n+1}=T-2\epsilon+t_{n+1}$.

1070: We claim that it is possible for the bureaucrat to go home by time $T$

1071: if and only if there exists a subset of $\{x_1,\ldots,x_n\}$ that sums to

1072: exactly~$T$.

1073: 

1074: If there is a subset of $\{x_1,\ldots,x_n\}$ that sums to exactly~$T$,

1075: then the bureaucrat can perform the corresponding subset of jobs (of

1076: total length $T$) and go home at time $T$; he is able to avoid doing

1077: any of the other jobs, since their critical times fall at an earlier

1078: time ($T-\epsilon$ or $T-2\epsilon$), making it infeasible to begin

1079: them at time $T$, by our assumption.

1080: 

1081: If, on the other hand, the bureaucrat is able to go home at time $T$,

1082: then we know the following:

1083: 

1084: \begin{enumerate}

1085: \item {\em The bureaucrat must have just completed a job at time

1086:     $T$.\/}

1087:   

1088:   He cannot quit a job and go home in the middle of a job, since the

1089:   job must have been completable at the instant he started (or

1090:   restarted) working on it, and it remains completable at the moment

1091:   that he would like to quit and go home.

1092: 

1093: \item {\em The bureaucrat must have been busy the entire time from 0

1094:     until time $T$.\/}

1095:   

1096:   He is not allowed to be idle for any period of time, since he could

1097:   always have been working on some available job, e.g., job $J_{n+1}$.

1098:   

1099: \item {\em If the bureaucrat starts a job, then he must finish it.\/}

1100:   

1101:   First, we note that if he starts job $J_i$ and does at least

1102:   $\epsilon$ of it, then he must finish it, since at time $T$ less

1103:   than $x_i-\epsilon$ remains to be done of the job, and it is not due

1104:   until time $T-\epsilon+x_i$, making it feasible to return to the job

1105:   at time $T$ (so that he cannot go home at time $T$).

1106:   

1107:   Second, we must consider the possibility that he may perform very

1108:   small amounts (less than $\epsilon$) of some jobs without finishing

1109:   them.  However, in this case, the {\em total\/} amount that he

1110:   completes of these barely started jobs is at most $n\epsilon\leq

1111:   {1\over 3}$.  This is a contradiction, since his total work time

1112:   consists of this fractional length of time, plus the sum of the

1113:   integral lengths of the jobs that he completed, which cannot add up

1114:   to the integer $T$.  Thus, in order for him to go home at exactly

1115:   time $T$, he must have completed every job that he started.

1116:   

1117:   Finally, note that he cannot use job $J_{n+1}$ as ``filler'', and do

1118:   part of it before going home at time $T$, since, if he starts it and

1119:   works at least time $2\epsilon$ on it, then, by the same reasoning

1120:   as above, he will be forced to stay and complete it.  Thus, he will

1121:   not start it at all, since he cannot complete it before time $T$

1122:   (recall that $t_{n+1}>T$).

1123: 

1124: \end{enumerate}

1125: %

1126: We conclude that the bureaucrat must complete a set of jobs whose lengths

1127: sum exactly to~$T$.

1128: 

1129: Thus, we have reduced {\sc Subset Sum} to our problem, showing that it is

1130: (weakly) NP-complete.

1131: 

1132: Note that the LBP we have constructed has non integer data. However,

1133: we can ``stretch'' time to get an equivalent problem in which all the

1134: data is integral. Letting $\epsilon = \frac{1}{3n}$, we multiply all job

1135: lengths and due dates by~$3n$.

1136: 

1137: \end{pf}

1138: 

1139: \begin{figure}[htbp]

1140: \begin{center}

1141: \input{fig2.pstex_t}

1142: \end{center}

1143: \caption{

1144:   Proof of hardness of LBP with preemption, assuming that all arrival

1145:   times are at time 0.}

1146: \label{fig:preempt-hard}

1147: \end{figure}

1148: 

1149: \noindent {\em Remark.}\quad 

1150: Hepner and Stein~\cite{HepnerStein2002} recently published a

1151: pseudo-polynomial-time algorithm for this problem, thus resolving an

1152: open problem from an earlier version of this

1153: paper~\cite{ArkinBenderMitchellSkiena99}.  

1154: \medskip

1155: 

1156: We come now to one of the main results of the paper.  We emphasize

1157: this result because it uses a rather sophisticated algorithm and

1158: analysis in order to show that, in contrast with the case of identical

1159: arrival times, the LBP with identical deadlines is polynomially

1160: solvable.  Specifically, the problems addressed in

1161: Theorems~\ref{thm:preempt-II.3}~and~\ref{thm:preempt-II.3-same-dead}

1162: are identical except that the flow of time is reversed.  Thus, we

1163: demonstrate that in contrast to most classical scheduling problems, in

1164: the LBP, when time flows in one direction, the problem is NP-hard,

1165: whereas when the flow of time is reversed, the problem is

1166: polynomial-time solvable.

1167: 

1168: The remainder of this section is devoted to proving the following

1169: theorem:

1170: 

1171: \begin{thm}

1172: \label{thm:preempt-II.3-same-dead}

1173: The LBP with preemption, under constraint~II (one can only work on

1174: jobs that can be completed) and objective~(3) (go home as early as

1175: possible), is solvable in polynomial time if all jobs have the same

1176: deadlines ($d_i=D$, for all $i$).  

1177: \end{thm}

1178: 

1179: We begin with a definition a ``forced gap:'' There is a {\em forced

1180:   gap\/} starting at time $\tau$ if $\tau$ is the earliest time such

1181: that the total work arriving by time $\tau$ is less than $\tau$.  This

1182: (first) forced gap ends at the arrival time, $\tau'$, of the next job.

1183: Subsequently, there may be more forced gaps, each determined by

1184: considering the scheduling problem that starts at the end, $\tau'$, of

1185: the previous forced gap.  We note that a forced gap can have length

1186: zero.

1187: 

1188: Under the ``go home early'' objective, we can assume, without loss of

1189: generality, that there are no forced gaps, since our problem really

1190: begins only at the time $\tau'$ that the {\em last\/} forced gap ends.

1191: (The bureaucrat is certainly not allowed to go home before the end

1192: $\tau'$ of the last forced gap, since more jobs arrive after $\tau'$

1193: that can be processed before their deadlines.)  While an optimal

1194: schedule may contain gaps that are not forced, the next lemma implies

1195: that there exists an optimal schedule having no unforced gaps.

1196: 

1197: \begin{lem}

1198:   Consider the LBP of Theorem~\ref{thm:preempt-II.3-same-dead}, and

1199:   assume that there are no forced gaps.  If there is a schedule having

1200:   makespan $T$, then there is a schedule with no gaps, also having

1201:   makespan~$T$.

1202: \end{lem}

1203: 

1204: \begin{pf}

1205:   Consider the first gap in the schedule, which begins at time $g$.

1206:   Because the gap is not forced, there is some job $j$ that is not

1207:   completed, and whose critical time is at time $g'\leq g$.  This is

1208:   because there must be a job that arrived before $g$ that is not

1209:   completed in the schedule, and at time $g$ it is no longer feasible

1210:   to complete it, and therefore its critical time is before $g$.  The

1211:   interval of time between $g'$ and $T$ may consist of (1) gaps, (2)

1212:   work on completed jobs, and (3) work on jobs that are never

1213:   completed. Consider a revised schedule in which, after time $g'$,

1214:   jobs of type~3 are removed, and jobs of type~2 are deferred to the

1215:   end of the schedule. (Since a job of type~2 is completed and all

1216:   jobs have the same deadline, we know that it is possible to move it

1217:   later in the schedule without passing its critical time.  It may not

1218:   be possible to move a (piece of a) job of type~3 later in the

1219:   schedule, since its critical time may have passed.)  In the revised

1220:   schedule, extend job~$j$ to fill the empty space. Note that there is

1221:   enough work in job $j$ to fill the space, since a critical time of

1222:   $g'$ means that the job must be executed continuously until deadline

1223:   $D$ in order to complete it.

1224: \end{pf}

1225: 

1226: \begin{lem} 

1227:   Consider an LBP of Theorem~\ref{thm:preempt-II.3-same-dead} in which

1228:   there are no forced gaps.  Any feasible schedule can be rearranged

1229:   so that all completed jobs are ordered by their arrival times and

1230:   all incomplete jobs are ordered by their arrival times.

1231: \end{lem}

1232: 

1233: \begin{pf}

1234:   The proof uses a simple exchange argument as in the standard proof

1235:   of optimality for the EDD (Earliest Due Date) policy in traditional

1236:   scheduling problems.

1237: \end{pf}

1238: 

1239: Our algorithm checks if there exists a schedule having no gaps that

1240: completes exactly at time $T$. Assume that the jobs $1,\ldots,n$ are

1241: labeled so that $a_1\leq a_2\cdots\leq a_n$.  The main steps of the

1242: algorithm are as follows:

1243: 

1244: \subsection*{The Algorithm:}

1245: 

1246: \begin{enumerate}

1247:   

1248: \item Determine the forced gaps.  This allows us to reduce to a

1249:   problem having no forced gaps, which starts at the end of the last

1250:   forced gap.

1251:   

1252:   The forced gaps are readily determined by computing the partial

1253:   sums, $\tau_j=\sum_{i=1}^j t_i$, for $j=0,1,\ldots,n$, and comparing

1254:   them to the arrival times.  (We define $\tau_0=0$.) The first forced

1255:   gap, then, begins at the time $\tau=\tau_{j^*}=\min\{\tau_j: \tau_j

1256:   < a_{j+1}\}$ and ends at time $a_{j^*+1}$.  ($\tau=0$ if $a_1>0$;

1257:   $\tau=\infty$ if there are no forced gaps.)  Subsequent forced gaps,

1258:   if any, are computed similarly, just by re-zeroing time at~$\tau'$,

1259:   and proceeding as with the first forced gap.

1260:   

1261: \item Let $x=D-T$ be the length of time between the common deadline

1262:   $D$ and our target makespan $T$.  A job $i$ for which $t_i\leq x$ is

1263:   called {\em short\/}; jobs for which $t_i> x$ are called {\em

1264:     long\/}.

1265:   

1266:   If it is {\em not\/} possible to schedule the set of short jobs so

1267:   that each is completed and they are all done by time $T$, then our

1268:   algorithm stops and returns ``NO,'' concluding that going home by

1269:   time $T$ is impossible.  Otherwise, we continue with the next step

1270:   of the algorithm.

1271:   

1272:   The rationale for this step is the observation that any job of

1273:   length at most $x$ must be completed in any schedule that permits

1274:   the bureaucrat to go home by time $T$, since its critical time

1275:   occurs at or after time~$T$.

1276:   

1277: \item Create a schedule ${\mathcal S}$ of all of the jobs, ordered by

1278:   their arrival times, in which the amount of time spent on job $i$ is

1279:   $t_i$ if the job is short (so it is done completely) and is $t_i-x$

1280:   if the job is long.

1281:   

1282:   For a long job $i$, $t_i-x$ is the maximum amount of time that can

1283:   be spent on this job without committing the bureaucrat to completing

1284:   the job, i.e., without causing the adjusted critical time of the job

1285:   to occur after time~$T$.

1286:   

1287:   If this schedule ${\mathcal S}$ has no gaps and ends at a time after

1288:   $T$, then our algorithm stops and returns ``YES.'' A feasible

1289:   schedule that allows the bureaucrat to go home by time $T$ is

1290:   readily constructed by ``squishing'' the schedule that we just

1291:   constructed: We reduce the amount of time spent on the long jobs,

1292:   starting with the latest long jobs and working backwards in time,

1293:   until the completion time of the last short job exactly equals~$T$.

1294:   This schedule completes all short jobs (as it should), and does

1295:   partial work on long jobs, leaving all of them with adjusted

1296:   critical times that fall {\em before\/} time $T$ (and are therefore

1297:   not possible to resume at time $T$, so they can be avoided).

1298:   

1299: \item If the above schedule ${\mathcal S}$ has gaps or ends before

1300:   time $T$, then ${\mathcal S}$ is not a feasible schedule for the

1301:   lazy bureaucrat, so we must continue the algorithm.

1302:   

1303:   Our objective is to decide {\em which\/} long jobs to complete, that

1304:   is, is there a set of long jobs to complete that will make it

1305:   possible to go home by time~$T$.  This problem is solved using the

1306:   dynamic programming algorithm {\em Schedule-by-$T$\/}, which is

1307:   described in detail below.

1308: \end{enumerate}

1309: 

1310: \subsection*{Procedure Schedule-by-{$T$}:}

1311: 

1312: Let $G_i$ be the sum of the gap lengths that occur before time $a_i$

1313: in schedule ${\mathcal S}$. Then, we know that in order to construct a

1314: gapless schedule, at least $\lceil G_i/x\rceil$ long jobs (in addition

1315: to the short jobs) from $1,\ldots,i-1$ must be completed. For each $i$

1316: we have such a constraint; collectively, we call these the {\em gap

1317:   constraints\/}.

1318: 

1319: \begin{claim}

1320:   If for each gap in schedule ${\mathcal S}$, there are enough long

1321:   jobs to be completed in order to fill the gap, then a feasible

1322:   schedule ending at $T$ exists.

1323: \end{claim}

1324: 

1325: We devise a dynamic programming algorithm as follows.  Let $T(m,k)$ be

1326: the earliest completion time of a schedule that satisfies the

1327: following:

1328: %

1329: \begin{enumerate}

1330: \item

1331: It completes by time $T$;

1332: 

1333: \item

1334: It uses jobs from the set $\{1,\ldots,k\}$;

1335: 

1336: \item

1337: It completes exactly $m$ jobs and does no other work (so it

1338: may have gaps, making it an infeasible schedule);

1339: 

1340: \item

1341: It satisfies the gap constraints; and

1342: 

1343: \item

1344: It completes all short jobs (of size $\leq x$).

1345: 

1346: \end{enumerate}

1347: %

1348: The boundary conditions on $T(m,k)$ are given by:

1349: %

1350: \begin{description}

1351: 

1352: \item[] $T(0,0)=0$; 

1353: 

1354: \item[] $T(0,n)=\infty$, which implies that at least one of the jobs

1355: must be completed;

1356: 

1357: \item[] $T(m,0)=\infty$ for $m>0$;

1358: 

1359: \item[] $T(m,k)=\infty$ if there exist constraints such that at least

1360: $m+1$ jobs from $1,\ldots,k$ must be completed, some of the jobs from

1361: $1,\ldots,k$ must be completed because they are short, and some

1362: additional jobs may need to be completed because of the gap

1363: constraints.  Note that this implies that $T(0,k)$ is equal to zero or

1364: infinity, depending on whether gap constraints are disobeyed.

1365: 

1366: \end{description}

1367: 

1368: In general, $T(m,k)$ is given by selecting the better of two options:

1369: %

1370: $$T(m,k)=\min\{\alpha ,\beta\},$$

1371: %

1372: where $\alpha$ is the earliest completion time if we choose not to

1373: execute job $k$ (which is a legal option only if job $k$ is long),

1374: giving

1375: %

1376: $$\alpha= \cases{T(m,k-1) & if  $t_k>x$ \cr

1377: \infty & otherwise,}$$

1378: %

1379: and $\beta$ is the earliest completion time if we choose to execute

1380: job $k$ (which is a legal option only if the resulting completion time

1381: is by time $T$), giving

1382: %

1383: $$\beta= \cases{\max (a_k+t_k,T(m-1,k-1)+t_k) & if this quantity is $\leq T$  \cr

1384: \infty & otherwise.}$$

1385: 

1386: \begin{lem}

1387:   There exists a feasible schedule completing at time $T$ if and only

1388:   if there exists an $m$ for which $T(m,n)<\infty$.

1389: \end{lem}

1390: 

1391: \begin{pf}

1392:   If $T(m,n)=\infty$ for all $m$, then, since the gap constraints

1393:   apply to any feasible schedule, and it is not possible to find such

1394:   a schedule for any number of jobs $m$, there is no feasible schedule

1395:   that completes on or before~$T$.

1396:   

1397:   If there exists an $m$ for which $T(m,n)<\infty$, let $m^*$ be the

1398:   smallest such $m$.  Then, by definition, $T(m^*,n)\leq T$. We show

1399:   that the schedule ${\mathcal S^*}$ obtained by the dynamic program

1400:   can be made into a feasible schedule ending at $T$. Consider jobs

1401:   that are not completed in the schedule ${\mathcal S}^*$; we wish to

1402:   use some of them to ``fill in'' the schedule to make it feasible, as

1403:   follows.

1404:   

1405:   Ordered by arrival times of incomplete jobs, and doing up to $t_i-x$

1406:   of each incomplete job, fill in the gaps. Note that by the gap

1407:   constraints, there is enough work to fill in all gaps. There are two

1408:   things that may make this schedule infeasible: (i) Some jobs are

1409:   worked on beyond their critical times, and (ii) the last job to be

1410:   done must be a completed one.

1411:   

1412:   (i). {\em Fixing the critical time problem:}\quad Consider a job $i$

1413:   that is processed at some time, beginning at $\tau$, after its

1414:   critical time, $c_i$.  We move all completed job pieces that fall

1415:   between $c_i$ and $T$ to the end of the schedule, lining them up to

1416:   end at $T$; then, we do job $i$ from time $c_i$ up until this batch

1417:   of completed jobs. This is legal because all completed jobs can be

1418:   pushed to the end of the schedule, and job $i$ cannot complete once

1419:   it stops processing.

1420:   

1421:   (ii). {\em Fixing the last job to be a complete one:}\quad Move a

1422:   ``sliver'' of the last completed job to just before time $T$. If

1423:   this is not possible (because the job would have to be done before

1424:   it arrives), then it means that we {\em must\/} complete one

1425:   additional job, so we consider $m^*+1$, and repeat the process.

1426:   

1427:   Note that a technical difficulty arises in the case in which the sum

1428:   of the gap lengths is an exact multiple of $x$: Do we have to

1429:   complete an additional job or not? This depends on whether we can

1430:   put a sliver of a completed job at $T$. There are several ways to

1431:   deal with this issue, including conditioning on the last completed

1432:   job, modifying the gap constraint, or ignoring the problem and

1433:   fixing it if it occurs (that is if we cannot put a sliver, then add

1434:   another job which must be completed to the gap constraint).

1435: \end{pf}

1436: 

1437: This completes the proof of our main theorem,

1438: Theorem~\ref{thm:preempt-II.3-same-dead}.

1439: 

1440: 

1441: \noindent {\em Remark.}\quad Even if all arrival times are the same,

1442: and deadlines are the same, and data is integer, an optimal solution

1443: may not be integer. In fact, there may not be an optimal solution,

1444: only a limiting one, as the following example shows: Let $a_i=0$ and

1445: $d_i=100$, for all $i$.  Jobs $1,\ldots,n-1$ have length 51, while job

1446: $n$ has length $t_{n}=48$.  A feasible schedule executes $\epsilon$ of

1447: each of the first $n-1 $ jobs, where $\epsilon >1/(n-2)$, and all of

1448: job~$n$, so the total work done is $(n -1) \epsilon

1449: +48>(n-1)/(n-2)+48$. Note that each of the first $n-1$ jobs have

1450: $51-\epsilon$ remaining to do, while there is only time

1451: $52-(n-1)\epsilon$ left before the deadline.  Now, by making

1452: $\epsilon$ arbitrarily close to $1/(n-2)$, we can make the schedule

1453: better and better.

1454: 

1455: 

1456: \section*{Acknowledgements}

1457: 

1458: We thank the referees for constructive comments and suggestions

1459: that improved the presentation of the paper.

1460: 

1461: \begin{thebibliography}{00}

1462: 

1463: \bibitem{ArkinBenderMitchellSkiena99}

1464: E.~M. Arkin, M.~A. Bender, J.~S.~B. Mitchell, and S.~S. Skiena.

1465: \newblock The Lazy Bureaucrat Scheduling Problem.

1466: \newblock {\em Proc.~6th Workshop on Discrete Algorithms (WADS)},

1467:   pages 80--85, 1999.

1468: 

1469: \bibitem{BGNS-99}

1470: A.~Bar-Noy, S.~Guha, J.~Naor, and B.~Schieber.

1471: \newblock Approximating the throughput of real-time multiple machine

1472:   scheduling.

1473: \newblock In {\em Proc.~31st ACM Symp. Theory of Computing (STOC)}, 

1474: pages 622--631, 1999.

1475: 

1476: \bibitem{BarvinokJWW1998}

1477: A.~ I.~Barvinok, D.~S.~Johnson, G.~J.~Woeginger, and R.~Woodroofe.

1478: \newblock The maximum traveling salesman problem under

1479: polyhedral norms. 

1480: \newblock In

1481: {\em Proc.~6th Conference on

1482: Integer Programming and Combinatorial Optimization (IPCO)\/},

1483: {\em Lecture Notes in Computer Science}, vol.~1412, 

1484: pages 195--201, 1998.

1485: 

1486: \bibitem{Fekete1999}

1487: S.~P.~Fekete.

1488: \newblock Simplicity and hardness of the maximum traveling 

1489: salesman problem under geometric distances. 

1490: \newblock {\em Proc.~10th {ACM-SIAM} 

1491: Symposium on Discrete Algorithms (SODA)}, pages

1492: 337--345, 1999.

1493: 

1494: \bibitem{GJ-79}

1495: M.~R. Garey and D.~S. Johnson.

1496: \newblock {\em Computers and Intractability: A Guide to the Theory of

1497:   {NP}-Completeness}.

1498: \newblock W. H. Freeman, San Francisco, 1979.

1499: 

1500: \bibitem{GoemansWilliamson1995}

1501: M.~X.~Goemans and D.~P.~Williamson.

1502: \newblock Improved approximation algorithms for maximum

1503: cut and satisfiability problems using semidefinite programming.

1504: \newblock {\em J. ACM}, 42:1115--1145, 1995.

1505: 

1506: \bibitem{HassinRubinstein1998}

1507: R.~Hassin and S.~Rubinstein. 

1508: \newblock An approximation algorithm for the maximum

1509: traveling salesman problem. 

1510: \newblock {\em Information Processing Letters}, 67(3):125-130, 1998.

1511: 

1512: \bibitem{HepnerStein2002}

1513: C.~Hepner and C.~Stein.

1514: \newblock Minimizing makespan for the lazy bureaucrat problem.

1515: \newblock In  {\em Proc.~8th

1516: Scandinavian Workshop on Algorithm Theory (SWAT)},

1517: pages 40--50, 2002.

1518: 

1519: \bibitem{KargerMotwaniRamkumar1997}

1520: D. Karger, R. Motwani, and G. Ramkumar. 

1521: \newblock On approximating the longest path in a graph. 

1522: \newblock {\em Algorithmica}, 18(1):82-98, 1997.

1523: 

1524: \bibitem{KargerSteinWein:1997}

1525: D.~Karger, C.~Stein, and J.~Wein.

1526: \newblock Scheduling algorithms.

1527: \newblock 

1528: In {\em Algorithms and Theory of Computation Handbook}, 

1529: CRC Press, 1999.

1530: 

1531: \bibitem{Keneally-82}

1532: T.~Keneally.

1533: \newblock {\em Schindler's List}.

1534: \newblock Touchstone Publishers, New York, 1993.

1535: 

1536: \bibitem{KosarajuParkStein1994}

1537: S.~R.~Kosaraju, J.~ K.~Park, and C.~Stein. 

1538: \newblock Long tours and short superstrings (preliminary version). 

1539: \newblock In {\em Proc.~35th Annual Symposium on Foundations of

1540: Computer Science (FOCS)}, pages 166--177, 1994. 

1541: 

1542: \bibitem{LLKS-93}

1543: E.~L.~Lawler, J.~K. Lenstra, A.~H.~G. Rinnooy Kan, and D.~B.~Shmoys.

1544: \newblock Sequencing and scheduling: Algorithms and complexity.

1545: \newblock S.C. Graves, P.H. Zipkin, and A.H.G. Rinnooy Kan (eds.) 

1546: \newblock In {\em Logistics of

1547:  Production and Inventory: Handbooks in Operations Research and Management Science},

1548:   volume~4, pages 445--522.  North-Holland, Amsterdam, 445-522, 1993.

1549: 

1550: \bibitem{LawlerMoore69}

1551: E.~L. Lawler and J.~M. Moore.

1552: \newblock A functional equation and its application to resource allocation and

1553:   sequencing problems.

1554: \newblock {\em Management Science}, 16:77--84, 1969.

1555: 

1556: \bibitem{Pinedo-95}

1557: M.~Pinedo.

1558: \newblock {\em Scheduling: {Theory}, Algorithms, and Systems}.

1559: \newblock Prentice Hall, 1995.

1560: 

1561: \end{thebibliography}

1562: 

1563: \end{document}

1564: 

1565: