Skip to main content
Log in

A best possible online algorithm for parallel batch scheduling with delivery times and limited restart

  • Original Paper
  • Published:
Optimization Letters Aims and scope Submit manuscript

Abstract

We consider the online scheduling on an unbounded (drop-line) parallel batch machine to minimize the time by which all jobs have been delivered. In this paper, all jobs arrive over time and the running batches are allowed limited restart. Here limited restart means that a running batch which contains restarted jobs cannot be restarted again. A drop-line parallel batch machine can process several jobs simultaneously as a batch, and all jobs in a batch start at the same time, and the completion time of a job equals the sum of its starting time and its processing time. Here we consider the restricted model: all jobs have agreeable processing times and delivery times. We provide a best possible online algorithm H with a competitive ratio of \(\frac{3}{2}\) for the problem on an unbounded parallel batch machine and the corresponding problem on an unbounded drop-line parallel batch machine, respectively.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Fig. 1
Fig. 2

Similar content being viewed by others

References

  1. Akker, M.V.D., Hoogeveen, H., Vakhania, N.: Restarts can help in the online minimization of the maximum delivery time on a single machine. J. Sched. 3, 333–341 (2003)

    Article  Google Scholar 

  2. Epstein, L., Van Stee, R.: Lower bounds for on-line single-machine scheduling. Theor. Comput. Sci. 299, 439–450 (2003)

    Article  MathSciNet  Google Scholar 

  3. Fu, R.Y., Tian, J., Yuan, J.J., He, C.: On-line scheduling on a batch machine to minimize makespan with limited restarts. Oper. Res. Lett. 36, 255–258 (2008)

    Article  MathSciNet  Google Scholar 

  4. Fu, R.Y., Cheng, T.C.E., Ng, C.T., Yuan, J.J.: Online scheduling on two parallel batching machines with limited restarts to minimize the makespan. Inf. Process. Lett. 110, 444–450 (2010)

    Article  MathSciNet  Google Scholar 

  5. Hoogeveen, H., Potts, C.N., Woeginger, G.J.: On-line Scheduling on a single machine: maximizing the number of early jobs. Oper. Res. Lett. 27, 193–197 (2000)

    Article  MathSciNet  Google Scholar 

  6. Hoogeveen, J.A., Vestjens, A.P.A.: A best possible deterministic on-line algorithm for minimizing maximum delivery time on a single machine. SIAM J. Discrete Math. 13, 56–63 (2000)

    Article  MathSciNet  Google Scholar 

  7. Lee, C.Y., Uzsoy, R., Martin-Vega, L.A.: Efficient algorithms for scheduling semi-conductor burn-in operations. Oper. Res. 40, 764–775 (1992)

    Article  MathSciNet  Google Scholar 

  8. Liu, P.H., Lu, X.W.: Online unbounded batch scheduling on parallel machines with delivery times. J. Comb. Optim. 29(1), 228–236 (2015)

    Article  MathSciNet  Google Scholar 

  9. Liu, H.L., Yuan, J.J., Li, W.J.: Online scheduling of equal length jobs on unbounded parallel batch processing machines with limited restart. J. Comb. Optim. 31, 1609–1622 (2016)

    Article  MathSciNet  Google Scholar 

  10. Tian, J., Cheng, T.C.E., Ng, C.T., Yuan, J.J.: An improved on-line algorithm for single parallel-batch machine scheduling with delivery times. Discrete Appl. Math. 160, 1191–1210 (2012)

    Article  MathSciNet  Google Scholar 

  11. Tian, J., Fu, R.Y., Yuan, J.J.: Online over time scheduling on parallel batch machines: a survey. J. Oper. Res. Soc. China 2, 445–454 (2014)

    Article  MathSciNet  Google Scholar 

  12. Tian, J., Wang, Q., Fu, R.Y., Yuan, J.J.: Online scheduling on the unbounded drop-line batch machines to minimize the maximum delivery completion time. Theor. Comput. Sci. 617, 65–68 (2016)

    Article  MathSciNet  Google Scholar 

  13. Uzsoy, R., Lee, C.Y., Martin-Vega, L.A.: A review of production planning and scheduling models in the semiconductor industry, part I: system characteristics, performance evaluation and production planning. IIE Trans. Sched. Logist. 24, 47–61 (1992)

    Google Scholar 

  14. Uzsoy, R., Lee, C.Y., Martin-Vega, L.A.: A survey of production planning and scheduling models in the semiconductor industry, part II: shop-floor control. IIE Trans. Sched. Logist. 26, 44–55 (1994)

    Google Scholar 

  15. Van Stee, R., Poutré, H.L.: Minimizing the total completion time on-line on a single machine, using restarts. J. Algorithms. 57, 95–129 (2005)

    Article  MathSciNet  Google Scholar 

  16. Wei, Z.G.: Scheduling on a batch machine with item-availability to minimize total weighted completion time. Master Degree Thesis, Zhengzhou University, (2011)

  17. Yuan, J.J., Li, S.S., Tian, J., Fu, R.Y.: A best on-line scheduling for the single machine parallel-batch scheduling with restricted delivery time. J. Comb. Optim. 17, 206–213 (2009)

    Article  MathSciNet  Google Scholar 

  18. Yuan, J.J., Fu, R.Y., Ng, C.T., Cheng, T.C.E.: A best online algorithm for unbounded parallel-batch scheduling with restarts to minimize makespan. J. Sched. 14, 361–369 (2011)

    Article  MathSciNet  Google Scholar 

Download references

Acknowledgements

This work was supported by NSFC (11701148, 11871213, 11571321,11501279), Henan University of Engineering (D2016017) and the Young Backbone Teachers of Henan Colleges (2019GGJS202).

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Xiwen Lu.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendix

Appendix

Proof of Lemma 2.4

Because \(S_n<S_{n-1}+p(B_{n-1})\), \(B_{n}\) is a restricted batch. Let \(p'\) be the longest processing time of jobs arriving in the time interval \((S_{n-1},S_{n}]\) and \(q''\) be the largest delivery time of jobs arriving in the time interval \((S_{n-1},S_{n}]\). Among the jobs which arrive in the time interval \((S_{n-1},S_{n}]\) and have the processing time \(p'\), select one which arrives latest as the job \(J'\). Let \(r'\) and \(q'\) be the arrival time and the delivery time of \(J'\), respectively. Among the jobs which arrive in the time interval \((S_{n-1},S_{n}]\) and have the delivery time \(q''\), select one which arrives latest as the job \(J''\). Let \(r''\) and \(q''\) be the arrival time and the delivery time of \(J''\), respectively. Thus we have \(p'\ge p''\) and \(q'\le q''\). Considering that both \(J'\) and \(J''\) arrive after the time \(S_{n-1}\), \(p'\ge p''\) and \(q'\le q''\), then by Lemma 2.1, we have \(L_{{\text{ opt }}}\ge S_{n-1}+p'+q''\). Since \(B_{n}\) is a restricted batch, by Step 4.2 and 4.3 of Algorithm H, we have \(p(B_{n-1})< p'<\frac{3}{2}p(B_{n-1})\) or \(\frac{3}{4}p(B_{n-1})< p'\le p(B_{n-1})\). Note that \(r'>S_{n-1}\ge \frac{1}{2}p(B_{n-1})\) by Observation 2.3.

Case 1.\(p(B_{n-1})< p'< \frac{3}{2}p(B_{n-1})\).

By Step 4.2 of Algorithm H, we have \(p(B_{n})=\max \{p(B_{n-1}), p'\}=p'\), \(S_{n}=\max \{p',r'\}\) and \(q(B_{n})=\max \{q'',q(B_{n-1})\}\).

As \(p(B_{n-1})<p'\), considering that all jobs have agreeable processing times and delivery times, then the largest delivery time of jobs in \(B_{n-1}\) is not more than \(q'\). Thus \(q(B_{n-1})\le q'\le q''\) and \(q(B_{n})=\max \{q'',q(B_{n-1})\}=q''\). Now \(L_{{\text{ on }}}=S_{n}+p(B_{n})+q(B_{n})=S_{n}+p'+q''\). Now we distinguish the following cases.

If \(r'>p'\), then \(S_{n}=r'\) and \(L_{{\text{ on }}}=S_{n}+p(B_{n})+q(B_{n})=r'+p'+q''\). Note that \(q'\le q''\). If \(q'=q''\), then \(L_{{\text{ opt }}}\ge r'+p'+q'=r'+p'+q''=L_{{\text{ on }}}\), which implies \(L_{{\text{ on }}}=L_{{\text{ opt }}}\). If \(q'<q''\), noting that all jobs have agreeable processing times and delivery times, then \(p'\le p''\). Also considering that \(p'\ge p''\), we have \(p'=p''\). By the definitions of \(J'\) and \(J''\) and \(p'=p''\), we have \(r'\ge r''\). Since \(B_{n}\) interrupts \(B_{n-1}\) at time \(S_{n}=r'\) and \(r'>p'\), then by Step 4.2 of Algorithm H, we have \(r'<\frac{5}{4}p(B_{n-1})\). Then \(r''\le r'<\frac{5}{4}p(B_{n-1})\). If \(r''<r'\) and \(r''\ge p''\), considering that \(p'=p''\) and \(r''<\frac{5}{4}p(B_{n-1})\), then by Step 4.2 of Algorithm H, \(B_{n}\) will interrupt \(B_{n-1}\) at time \(r''\), a contradiction to \(S_{n}=r'\) and \(r''<r'\). If \(r''<r'\) and \(r''<p''\), considering that \(p''=p'\), \(r'>p'\) and \(p'\) is the longest processing time of jobs arriving in the time interval \((S_{n-1},S_{n}]=(S_{n-1},r']\), then by Step 4.2 of Algorithm H, \(B_{n}\) will interrupt \(B_{n-1}\) at time \(p''=p'\), a contradiction to the fact that \(S_{n}=r'\) and \(r'>p'\). So the assumption of \(r''<r'\) is not right and we have \(r''=r'\). So if \(q'<q''\), we have \(p'=p''\) and \(r''=r'\). Then \(L_{{\text{ opt }}}\ge r''+p''+q''=r'+p'+q''=L_{{\text{ on }}}\), which implies \(L_{{\text{ on }}}=L_{{\text{ opt }}}\).

If \(r'\le p'\), then \(S_{n}=p'\) and \(L_{{\text{ on }}}=S_{n}+p(B_{n})+q(B_{n})=2p'+q''\). Thus \(\frac{L_{{\text{ on }}}}{L_{{\text{ opt }}}}\le \frac{2p'+q''}{S_{n-1}+p'+q''}\le \frac{2p'+q''}{\frac{1}{2}p(B_{n-1})+p'+q''}\le \frac{2p'}{\frac{1}{2}p(B_{n-1})+p'}\le \frac{2\times \frac{3}{2}p(B_{n-1})}{\frac{3}{2}p(B_{n-1})+\frac{1}{2}p(B_{n-1})}=\frac{3}{2}\).

Case 2.\(\frac{3}{4}p(B_{n-1})< p'\le p(B_{n-1})\).

By Step 4.3 of Algorithm H, we have \(p(B_{n})=\max \{p',p(B_{n-1})\}=p(B_{n-1})\), \(r'<\frac{5}{4}p(B_{n-1})\) and \(S_{n}=\max \{2p(B_{n-1})-p',r'\}\). Note that \(S_{n}< \frac{5}{4}p(B_{n-1})\) as \(\frac{3}{4}p(B_{n-1})< p'\) and \(r'<\frac{5}{4}p(B_{n-1})\). Then \(L_{{\text{ on }}}=S_{n}+p(B_{n})+q(B_{n})=S_{n}+p(B_{n-1})+q(B_{n})<\frac{9}{4}p(B_{n-1})+q(B_{n})\).

Subcase 2.1\(q''>q(B_{n-1})\).

Then \(q(B_{n})=\max \{q(B_{n-1}),q''\}=q''\). Now \(L_{{\text{ on }}}=S_{n}+p(B_{n})+q(B_{n})=S_{n}+p(B_{n-1})+q''<\frac{9}{4}p(B_{n-1})+q''\).

Note that \(q_{n-1}^{p}\) is the delivery time of \(J_{n-1}^{p}\) and \(p(B_{n-1})\) is the processing time of \(J_{n-1}^{p}\). Considering that \(q''>q(B_{n-1})\ge q_{n-1}^{p}\), we have \(p''\ge p(B_{n-1})\). Considering that \(p''\le p'\), \(p'\le p(B_{n-1})\) and \(p''\ge p(B_{n-1})\), we have \(p''=p'=p(B_{n-1})\). Then \(L_{{\text{ opt }}}\ge S_{n-1}+p'+q''\ge \frac{1}{2}p(B_{n-1})+p(B_{n-1})+q''=\frac{3}{2}p(B_{n-1})+q''\) by Observation 2.3. Thus \(\frac{L_{{\text{ on }}}}{L_{{\text{ opt }}}}\le \frac{\frac{9}{4}p(B_{n-1})+q''}{\frac{3}{2}p(B_{n-1})+q''}\le \frac{3}{2}\).

Subcase 2.2\(q''\le q(B_{n-1})\).

Then \(q(B_{n})=\max \{q(B_{n-1}),q''\}=q(B_{n-1})\). Now \(L_{{\text{ on }}}=S_{n}+p(B_{n})+q(B_{n})=S_{n}+p(B_{n-1})+q(B_{n-1})\). Since \(B_{n}\) interrupts \(B_{n-1}\), we have \(p'\ge \frac{3}{4}p(B_{n-1})+\frac{1}{4}q(B_{n-1})\) by Step 4.3 of Algorithm H. As \(p'\ge \frac{3}{4}p(B_{n-1})+\frac{1}{4}q(B_{n-1})\) and \(p'\le p(B_{n-1})\), we have \(q(B_{n-1})\le p(B_{n-1})\).

By Lemma 2.2, for the batch \(B_{n-1}\), at least one of the jobs \(J_{n-1}^{p}\) and \(J_{n-1}^{q}\) has the processing time \(p(B_{n-1})\) and the delivery time \(q(B_{n-1})\) and denote the job by \(J_{n-1}^{*}\).

If \(S_{n}=2p(B_{n-1})-p'\), then \(L_{{\text{ on }}}=S_{n}+p(B_{n-1})+q(B_{n-1})=2p(B_{n-1})-p'+p(B_{n-1})+q(B_{n-1})\le 2p(B_{n-1})-(\frac{3}{4}p(B_{n-1})+\frac{1}{4}q(B_{n-1}))+p(B_{n-1})+q(B_{n-1})= \frac{9}{4}p(B_{n-1})+\frac{3}{4}q(B_{n-1})\). Note that \(J'\) arrives after time \(S_{n-1}\), and the processing time of \(J_{n-1}^{*}\) is \(p(B_{n-1})\), and the delivery time of \(J_{n-1}^{*}\) is \(q(B_{n-1})\). Thus \(L_{{\text{ opt }}}\ge \min \{S_{n-1}+p(B_{n-1})+q(B_{n-1}),p(B_{n-1})+p'\}\) whether \(J'\) and \(J_{n-1}^{*}\) are in a common batch in the optimal off-line schedule \(\pi \) or not. Then \(L_{{\text{ opt }}}\ge \min \{S_{n-1}+p(B_{n-1})+q(B_{n-1}),p(B_{n-1})+p'\}\ge \min \{\frac{1}{2}p(B_{n-1})+p(B_{n-1})+q(B_{n-1}),p(B_{n-1})+\frac{3}{4}p(B_{n-1})+\frac{1}{4}q(B_{n-1})\}\ge \min \{\frac{3}{2}p(B_{n-1})+q(B_{n-1}),p(B_{n-1})+\frac{2}{4}p(B_{n-1})+\frac{2}{4}q(B_{n-1})\}=\frac{3}{2}p(B_{n-1})+\frac{1}{2}q(B_{n-1})\). Here the last but one inequality hods as \(p(B_{n-1})\ge q(B_{n-1})\). Thus \(\frac{L_{{\text{ on }}}}{L_{{\text{ opt }}}}\le \frac{\frac{9}{4}p(B_{n-1})+\frac{3}{4}q(B_{n-1})}{\frac{3}{2}p(B_{n-1})+\frac{1}{2}q(B_{n-1})}=\frac{3}{2}\).

If \(S_{n}=r'\), then \(r'\ge 2p(B_{n-1})-p'\) and \(L_{{\text{ opt }}}\ge r'+p'\ge 2p(B_{n-1})\). Now \(L_{{\text{ on }}}=S_{n}+p(B_{n-1})+q(B_{n-1})=r'+p(B_{n-1})+q(B_{n-1})\). Considering that \(p'\ge \frac{3}{4}p(B_{n-1})+\frac{1}{4}q(B_{n-1})\) and \(q(B_{n-1})\le p(B_{n-1})\), we have \(\frac{L_{{\text{ on }}}}{L_{{\text{ opt }}}}\le \frac{r'+p(B_{n-1})+q(B_{n-1})}{r'+p'}=1+\frac{p(B_{n-1})+q(B_{n-1})-p'}{r'+p'}\le 1+\frac{\frac{1}{4}p(B_{n-1})+\frac{3}{4}q(B_{n-1})}{2p(B_{n-1})}\le 1+\frac{p(B_{n-1})}{2p(B_{n-1})}=\frac{3}{2}\). The result follows. \(\square \)

Proof of Lemma 2.5

As \(S_n=S_{n-1}+p(B_{n-1})\), then \(B_{n}\) is a free batch and all jobs in \(B_{n}\) arrive after the time \(S_{n-1}\). Now \(L_{{\text{ on }}}=S_{n}+p(B_{n})+q(B_{n})=S_{n-1}+p(B_{n-1})+p(B_{n})+q(B_{n})\). By Lemma 2.2, \(L_{{\text{ opt }}}\ge r(B_{n})+p(B_n)+q(B_n)>S_{n-1}+p(B_n)+q(B_n)\). Thus

$$\begin{aligned} L_{{\text{ on }}}-L_{{\text{ opt }}}\le p(B_{n-1}). \end{aligned}$$
(1)

Let \(p'\) be the longest processing time of jobs arriving in the time interval \((S_{n-2},S_{n-1}]\). Among the jobs which arrive in the time interval \((S_{n-2},S_{n-1}]\) and have a processing time \(p'\), select one which arrives latest as the job \(J'\). Let \(r'\) and \(q'\) be the arrival time and the delivery time of \(J'\), respectively. Considering that \(B_{n-1}\) is a restricted batch, then by Algorithm H, we have \(p(B_{n-2})< p'<\frac{3}{2}p(B_{n-2})\) or \(\frac{3}{4}p(B_{n-2})<p'\le p(B_{n-2})\). By Lemma 2.2, at least one of the jobs \(J_{n}^{p}\) and \(J_{n}^{q}\) has the processing time of \(p(B_{n})\) and the delivery time of \(q(B_{n})\). Let \(J_{n}^{*}\) be one of \(J_{n}^{p}\) and \(J_{n}^{q}\) whose processing time is \(p(B_{n})\) and delivery time is \(q(B_{n})\).

Case 1.\(p(B_{n-2})< p'<\frac{3}{2}p(B_{n-2})\).

Then \(p(B_{n-1})=\max \{p(B_{n-2}),p'\}=p'\). By Step 4.2 of Algorithm H, \(S_{n-1}=\max \{p',r'\}\).

If \(r'\ge p'\), we have \(S_{n-1}=r'\) and \(L_{{\text{ opt }}}\ge r'+p'\ge 2p'=2p(B_{n-1})\). Then by the inequality (1), we have \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{p(B_{n-1})}{2p(B_{n-1})}=\frac{1}{2}\).

If \(r'<p'\), we have \(S_{n-1}=p'=p(B_{n-1})\). If \(J'\) starts not before \(J_{n}^{*}\) in the optimal off-line schedule \(\pi \), considering that \(J_{n}^{*}\) arrives after the time \(S_{n-1}\), then \(L_{{\text{ opt }}}\ge S_{n-1}+p'=2p(B_{n-1})\). Thus we have \(L_{{\text{ on }}}-L_{{\text{ opt }}}\le p(B_{n-1})\le \frac{1}{2}L_{{\text{ opt }}}\). If \(J'\) starts before \(J_{n}^{*}\) in the schedule \(\pi \), then \(L_{{\text{ opt }}}\ge r'+p'+p(B_{n})+q(B_{n})>S_{n-2}+p(B_{n-1})+p(B_{n})+q(B_{n})\). Considering that \(\frac{1}{3}p(B_{n-1})=\frac{1}{3}p'<\frac{1}{2}p(B_{n-2})\) and \(S_{n-2}\ge \frac{1}{2}p(B_{n-2})\), we have \(S_{n-2}\ge \frac{1}{3}p(B_{n-1})\). Then \(L_{{\text{ opt }}}>S_{n-2}+p(B_{n-1})+p(B_{n})+q(B_{n})\ge \frac{1}{3}p(B_{n-1})+p(B_{n-1})=\frac{4}{3}p(B_{n-1})\) and \(L_{{\text{ on }}}-L_{{\text{ opt }}}\le S_{n-1}+p(B_{n-1})+p(B_{n})+q(B_{n})-(S_{n-2}+p(B_{n-1})+p(B_{n})+q(B_{n}))=S_{n-1}-S_{n-2}\le p(B_{n-1})-\frac{1}{3}p(B_{n-1})=\frac{2}{3}p(B_{n-1})\). Thus \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{\frac{2}{3}p(B_{n-1})}{\frac{4}{3}p(B_{n-1})}=\frac{1}{2}\).

Case 2.\(\frac{3}{4}p(B_{n-2})< p'\le p(B_{n-2})\).

Then \(p(B_{n-1})=\max \{p(B_{n-2}),p'\}=p(B_{n-2})\). By Step 4.3 of Algorithm H, we have \(S_{n-1}=\max \{2p(B_{n-2})-p',r'\}\). Then \(L_{{\text{ on }}}=S_{n-1}+p(B_{n-1})+p(B_{n})+q(B_{n})=S_{n-1}+p(B_{n-2})+p(B_{n})+q(B_{n})\).

If \(r'\ge 2p(B_{n-2})-p'\), then \(L_{{\text{ opt }}}\ge r'+p'\ge 2p(B_{n-2})=2p(B_{n-1})\). Thus \(L_{{\text{ on }}}-L_{{\text{ opt }}}\le p(B_{n-1})\le \frac{1}{2}L_{{\text{ opt }}}\) by the inequality (1).

If \(r'<2p(B_{n-2})-p'\), then \(S_{n-1}=2p(B_{n-2})-p'<\frac{5}{4}p(B_{n-2})\). Now we distinguish the following cases. If \(J_{n-2}^{p}\) or \(J'\) starts not before \(J_{n}^{*}\) in the schedule \(\pi \), considering that \(J_{n}^{*}\) arrives after the time \(S_{n-1}\), then \(L_{{\text{ opt }}}\ge S_{n-1}+\min \{p(B_{n-2}),p'\}=S_{n-1}+p'=2p(B_{n-2})=2p(B_{n-1})\). Thus \(L_{{\text{ on }}}-L_{{\text{ opt }}}\le p(B_{n-1})\le \frac{1}{2}L_{{\text{ opt }}}\). If both \(J_{n-2}^{p}\) and \(J'\) start before \(J_{n}^{*}\) in the schedule \(\pi \), then \(J_{n}^{*}\) starts not earlier than the time \(\min \{p(B_{n-2})+p',r'+p(B_{n-2})\}\) in the schedule \(\pi \) whether \(J_{n-2}^{p}\) and \(J'\) are in a common batch or not in the schedule \(\pi \). Note that \(\min \{p(B_{n-2})+p',r'+p(B_{n-2})\}\ge \min \{p(B_{n-2})+\frac{3}{4}p(B_{n-2}),S_{n-2}+p(B_{n-2})\}\ge \min \{p(B_{n-2})+\frac{3}{4}p(B_{n-2}),\frac{1}{2}p(B_{n-2})+p(B_{n-2})\}= \frac{3}{2}p(B_{n-2})\). Then \(J_{n}^{*}\) starts not earlier than the time \(\frac{3}{2}p(B_{n-2})\) in the schedule \(\pi \) and \(L_{{\text{ opt }}}\ge \frac{3}{2}p(B_{n-2})+p(B_{n})+q(B_{n})\). Thus \(\frac{L_{{\text{ on }}}}{L_{{\text{ opt }}}}\le \frac{S_{n-1}+p(B_{n-2})+p(B_{n})+q(B_{n})}{\frac{3}{2}p(B_{n-2})+p(B_{n})+q(B_{n})}\le \frac{\frac{5}{4}p(B_{n-2})+p(B_{n-2})+p(B_{n})+q(B_{n})}{\frac{3}{2}p(B_{n-2})+p(B_{n})+q(B_{n})}\le \frac{3}{2}\). The result follows. \(\square \)

Proof of Lemma 2.6

Since \(S_n=S_{n-1}+p(B_{n-1})\), then \(B_{n}\) is a free batch, and \(r(B_{n})>S_{n-1}\), and \(L_{{\text{ on }}}=S_{n-1}+p(B_{n-1})+p(B_{n})+q(B_{n})\). By Lemma 2.2, \(L_{{\text{ opt }}}\ge r(B_{n})+p(B_{n})+q(B_{n})>S_{n-1}+p(B_{n})+q(B_{n})\). Then

$$\begin{aligned} L_{{\text{ on }}}-L_{{\text{ opt }}}\le p(B_{n-1}). \end{aligned}$$
(2)

Case 1.\(p(B_{n})\ge \frac{3}{2}p(B_{n-1})\).

Note that \(L_{{\text{ on }}}-L_{{\text{ opt }}}\le p(B_{n-1})\) and \(S_{n-1}\ge \frac{1}{2}p(B_{n-1})\) by Observation 2.3. Then \(L_{{\text{ opt }}}>S_{n-1}+p(B_{n})+q(B_{n})\ge \frac{1}{2}p(B_{n-1})+\frac{3}{2}p(B_{n-1})=2p(B_{n-1})\). Thus \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{p(B_{n-1})}{2p(B_{n-1})}=\frac{1}{2}\).

Case 2.\(p(B_{n-1})< p(B_{n})< \frac{3}{2}p(B_{n-1})\).

Note that the arrival time of \(J_{n}^{p}\) is \(r_{n}^{p}\) and the processing time of \(J_{n}^{p}\) is \(p(B_{n})\). Now we prove that \(r_{n}^{p}\ge p(B_{n-1})\). If \(r_{n}^{p}<p(B_{n-1})\), noting that \(r_{n}^{p}<p(B_{n-1})<p(B_{n})\), then \(B_{n}\) will interrupt \(B_{n-1}\) at time \(p(B_{n})\) by Step 4.2 of Algorithm H, a contradiction. So we have \(r_{n}^{p}\ge p(B_{n-1})\). Then \(L_{{\text{ opt }}}\ge r_{n}^{p}+p(B_{n})\ge 2p(B_{n-1})\). Thus by the inequality (2), \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{p(B_{n-1})}{2p(B_{n-1})}=\frac{1}{2}\).

Case 3.\(p(B_{n})\le p(B_{n-1})\).

If \(r_{n}^{p}+p(B_{n})\ge 2p(B_{n-1})\), we have \(L_{{\text{ opt }}}\ge r_{n}^{p}+p(B_{n})\ge 2p(B_{n-1})\). Then \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{p(B_{n-1})}{2p(B_{n-1})}=\frac{1}{2}\). Now we assume that \(r_{n}^{p}+p(B_{n})<2p(B_{n-1})\), i.e., \(r_{n}^{p}< 2p(B_{n-1})-p(B_{n})\).

Subcase 3.1.\(p(B_{n})=p(B_{n-1})\).

If \(p(B_{n-1})>q(B_{n-1})\), then \(\frac{3}{4}p(B_{n-1})+\frac{1}{4}q(B_{n-1})<p(B_{n-1})=p(B_{n})\). Considering that \(r_{n}^{p}<2p(B_{n-1})-p(B_{n})=p(B_{n-1})=p(B_{n})\) and \(p(B_{n})>\frac{3}{4}p(B_{n-1})+\frac{1}{4}q(B_{n-1})\), then \(B_{n}\) will interrupt \(B_{n-1}\) at time \(2p(B_{n-1})-p(B_{n})=p(B_{n-1})\) by Step 4.3 of Algorithm H, a contradiction. So we have \(p(B_{n-1})\le q(B_{n-1})\). By Lemma 2.2, \(L_{{\text{ opt }}}\ge p(B_{n-1})+q(B_{n-1})\ge 2p(B_{n-1})\). Then by the inequality (2), we have \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{p(B_{n-1})}{2p(B_{n-1})}=\frac{1}{2}\).

Subcase 3.2.\(p(B_{n})<p(B_{n-1})\).

By Lemma 2.2, for the batch \(B_{n-1}\), at least one of the jobs \(J_{n-1}^{p}\) and \(J_{n-1}^{q}\) has the processing time of \(p(B_{n-1})\) and the delivery time of \(q(B_{n-1})\). So let \(J_{n-1}^{*}\) be one of the jobs \(J_{n-1}^{p}\) and \(J_{n-1}^{q}\) whose processing time is \(p(B_{n-1})\) and delivery time is \(q(B_{n-1})\). Similarly, let \(J_{n}^{*}\) be one of the jobs \(J_{n}^{p}\) and \(J_{n}^{q}\) whose processing time is \(p(B_{n})\) and delivery time is \(q(B_{n})\).

If \(p(B_{n-1})\le q(B_{n-1})\), by Lemma 2.2, we have \(L_{{\text{ opt }}}\ge p(B_{n-1})+q(B_{n-1})\ge 2p(B_{n-1})\). Then \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{p(B_{n-1})}{2p(B_{n-1})}=\frac{1}{2}\). Now we assume that \(p(B_{n-1})>q(B_{n-1})\).

If \(\frac{3}{4}p(B_{n-1})+\frac{1}{4}q(B_{n-1})<p(B_{n})<p(B_{n-1})\), considering that \(r_{n}^{p}< 2p(B_{n-1})-p(B_{n})<\frac{5}{4}p(B_{n-1})\), then \(B_{n}\) will interrupt \(B_{n-1}\) by Step 4.3 of Algorithm H, a contradiction. So we have \(p(B_{n})\le \frac{3}{4}p(B_{n-1})+\frac{1}{4}q(B_{n-1})\).

If \(J_{n-1}^{*}\) starts not earlier than the time \(S_{n-1}\) in the schedule \(\pi \), then \(L_{{\text{ opt }}}\ge S_{n-1}+p(B_{n-1})+q(B_{n-1})\). Note that the processing time of \(J_{n}^{q}\) is \(p_{n}^{q}\) and the processing time of \(J_{n-1}^{p}\) is \(p(B_{n-1})\). Then \(p_{n}^{q}\le p(B_{n})<p(B_{n-1})\). Considering that \(p_{n}^{q}<p(B_{n-1})\), then the delivery time of \(J_{n}^{q}\) is not more than the delivery time of \(J_{n-1}^{p}\), i.e., \(q(B_{n})\le q_{n-1}^{p}\). Then \(q(B_{n})\le q_{n-1}^{p}\le q(B_{n-1})\). Thus \(L_{{\text{ on }}}-L_{{\text{ opt }}}\le S_{n-1}+p(B_{n-1})+p(B_{n})+q(B_{n})-(S_{n-1}+p(B_{n-1})+q(B_{n-1}))\le p(B_{n})\le \frac{3}{4}p(B_{n-1})+\frac{1}{4}q(B_{n-1})\). Note that \(L_{{\text{ opt }}}\ge S_{n-1}+p(B_{n-1})+q(B_{n-1})\ge \frac{1}{2}p(B_{n-1})+p(B_{n-1})+q(B_{n-1})\). Then \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{1}{2}\).

Now we assume that \(J_{n-1}^{*}\) starts earlier than the time \(S_{n-1}\) in the schedule \(\pi \). Considering that all jobs in \(B_{n}\) arrive after the time \(S_{n-1}\), then \(J_{n-1}^{*}\) starts before \(J_{n}^{*}\) in the schedule \(\pi \). Then \(L_{{\text{ opt }}}\ge r(B_{n-1})+p(B_{n-1})+p(B_{n})+q(B_{n})\).

As \(B_{n-1}\) is a free batch, by Observation 2.1, we have \(S_{n-1}=\frac{1}{2}p(B_{n-1})\), or \(S_{n-1}=r(B_{n-1})\), or \(S_{n-1}=S_{n-2}+p(B_{n-2})\).

If \(S_{n-1}=r(B_{n-1})\) or \(S_{n-1}=\frac{1}{2}p(B_{n-1})\), then \(L_{{\text{ on }}}-L_{{\text{ opt }}}\le S_{n-1}+p(B_{n-1})+p(B_{n})+q(B_{n})-(r(B_{n-1})+p(B_{n-1})+p(B_{n})+q(B_{n}))=S_{n-1}-r(B_{n-1})\le \frac{1}{2}p(B_{n-1})\) and \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{1}{2}\).

From now on we assume that \(S_{n-1}=S_{n-2}+p(B_{n-2})\). So \(L_{{\text{ on }}}=S_{n-2}+p(B_{n-2})+p(B_{n-1})+p(B_{n})+q(B_{n})\). Since \(B_{n-1}\) is a free batch, we have \(r(B_{n-1})>S_{n-2}\). Thus \(L_{{\text{ opt }}}\ge r(B_{n-1})+p(B_{n-1})+p(B_{n})+q(B_{n})>S_{n-2}+p(B_{n-1})+p(B_{n})+q(B_{n})\) and \(L_{{\text{ on }}}-L_{{\text{ opt }}}\le p(B_{n-2})\). Noting that \(L_{{\text{ on }}}-L_{{\text{ opt }}}\le p(B_{n-1})\), then

$$\begin{aligned} L_{{\text{ on }}}-L_{{\text{ opt }}}\le \min \{p(B_{n-1}),p(B_{n-2})\}. \end{aligned}$$
(3)

By Lemma 2.2, one job of the batch \(B_{n-2}\) has the processing time of \(p(B_{n-2})\) and the delivery time of \(q(B_{n-2})\). Let \(J_{n-2}^{*}\) be such a job in \(B_{n-2}\) with the processing time \(p(B_{n-2})\) and the delivery time \(q(B_{n-2})\).

If \(B_{n-2}\) is a restricted batch, then we have \(S_{n-2}\ge p(B_{n-2})\) by Observation 2.2. Thus \(L_{{\text{ opt }}}\ge r(B_{n})>S_{n-1}=S_{n-2}+p(B_{n-2})\ge 2p(B_{n-2})\). Hence \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}} \le \frac{\min \{p(B_{n-1}),p(B_{n-2})\}}{2p(B_{n-2})}\le \frac{1}{2}\) by the inequality (3).

If \(B_{n-2}\) is a free batch, we distinguish the following cases. If \(p(B_{n-1})\ge \frac{3}{2}p(B_{n-2})\), we have \(L_{{\text{ opt }}}\ge r(B_{n-1})+p(B_{n-1})>S_{n-2}+p(B_{n-1})\ge \frac{1}{2}p(B_{n-2})+\frac{3}{2}p(B_{n-2})=2p(B_{n-2})\). Thus \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{\min \{p(B_{n-1}),p(B_{n-2})\}}{2p(B_{n-2})}\le \frac{1}{2}\). If \(p(B_{n-2})<p(B_{n-1})< \frac{3}{2}p(B_{n-2})\) or [\(\frac{3}{4}p(B_{n-2})<p(B_{n-1})\le p(B_{n-2})\) and \(p(B_{n-1})\ge \frac{3}{4}p(B_{n-2})+\frac{1}{4}q(B_{n-2})\)], noting that \(B_{n-1}\) does not interrupt the free batch \(B_{n-2}\), then \(r_{n-1}^{p}\ge \frac{5}{4}p(B_{n-2})\) by Step 4.2 and 4.3 of Algorithm H. Note that the arrival time of \(J_{n-1}^{p}\) is \(r_{n-1}^{p}\) and the processing time of \(J_{n-1}^{p}\) is \(p(B_{n-1})\). Thus \(L_{{\text{ opt }}}\ge r_{n-1}^{p}+p(B_{n-1})\ge \frac{5}{4}p(B_{n-2})+\frac{3}{4}p(B_{n-2})\ge 2\min \{p(B_{n-1}),p(B_{n-2})\}\). So \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{\min \{p(B_{n-1}),p(B_{n-2})\}}{2\min \{p(B_{n-1}),p(B_{n-2})\}}\le \frac{1}{2}\). If [\(\frac{3}{4}p(B_{n-2})<p(B_{n-1})\le p(B_{n-2})\) and \(p(B_{n-1})<\frac{3}{4}p(B_{n-2})+\frac{1}{4}q(B_{n-2})\)] or \(p(B_{n-1})\le \frac{3}{4}p(B_{n-2})\), we distinguish the following two cases. If \(J_{n-1}^{p}\) and \(J_{n-2}^{*}\) are not in a common batch in the schedule \(\pi \), \(L_{{\text{ opt }}}\ge p(B_{n-1})+p(B_{n-2})\ge 2\min \{p(B_{n-1}),p(B_{n-2})\}\). Thus \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{\min \{p(B_{n-1}),p(B_{n-2})\}}{2\min \{p(B_{n-1}),p(B_{n-2})\}}=\frac{1}{2}\). If \(J_{n-1}^{p}\) and \(J_{n-2}^{*}\) are in a common batch in the schedule \(\pi \), then \(L_{{\text{ opt }}}\ge r_{n-1}^{p}+p(B_{n-2})+q(B_{n-2})>S_{n-2}+p(B_{n-2})+q(B_{n-2})\ge \frac{1}{2}p(B_{n-2})+p(B_{n-2})+q(B_{n-2})= \frac{3}{2}p(B_{n-2})+q(B_{n-2})\). Thus \(\frac{L_{{\text{ on }}}-L_{{\text{ opt }}}}{L_{{\text{ opt }}}}\le \frac{p(B_{n-1})}{\frac{3}{2}p(B_{n-2})+q(B_{n-2})}\le \frac{\frac{3}{4}p(B_{n-2})+\frac{1}{4}q(B_{n-2})}{\frac{3}{2}p(B_{n-2})+q(B_{n-2})}\le \frac{1}{2}\). The result follows. \(\square \)

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Liu, H., Lu, X. & Li, W. A best possible online algorithm for parallel batch scheduling with delivery times and limited restart. Optim Lett 15, 1155–1173 (2021). https://doi.org/10.1007/s11590-020-01618-8

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s11590-020-01618-8

Keywords

Navigation