## Queueing Disciplines

Venue: | WILEY ENCYCLOPEDIA OF OPERATIONS RESEARCH AND MANAGEMENT SCIENCE |

Citations: | 1 - 0 self |

### BibTeX

@MISC{Harchol-Balter_queueingdisciplines,

author = {Mor Harchol-Balter},

title = {Queueing Disciplines },

year = {}

}

### OpenURL

### Abstract

### Citations

1477 | Wide-Area Traffic: The Failure of Poisson Modeling
- Paxson, Floyd
- 1995
(Show Context)
Citation Context ... [1, 2]) (where C2 values of 25 to 49 have been measured), sizes of files transferred through the Web [3, 4], sizes of files stored in Unix filesystems [5], durations of FTP transfers in the Internet =-=[6]-=-, and CPU requirements for supercomputing jobs [7]. In this article we will therefore focus a lot of attention on the effect of job size variability. Decreasing failure rate (hereafter referred to as ... |

1178 | Self-similarity in World Wide Web traffic: evidence and possible causes
- Crovella, Bestavros
- 1996
(Show Context)
Citation Context ...LogNormal, or a Weibull with high C 2 X . Some examples include Unix process CPU requirements (see [1, 2]) (where C2 values of 25 to 49 have been measured), sizes of files transferred through the Web =-=[3, 4]-=-, sizes of files stored in Unix filesystems [5], durations of FTP transfers in the Internet [6], and CPU requirements for supercomputing jobs [7]. In this article we will therefore focus a lot of atte... |

651 |
Stochastic Processes
- Ross
- 1983
(Show Context)
Citation Context ... only true when jobs have (4) 5deterministic job sizes. When job sizes are variable, a random arrival is more likely to see a larger-than-average excess, and, in fact, by the Inspection Paradox (see =-=[16]-=-), the expected excess turns out to be E [ X 2] /2E[X], which comes up in the derivation of formula (4) above. From the mean waiting time, it is easy to derive the mean response time as follows: E [T(... |

320 | Exploiting process lifetime distributions for dynamic load balancing
- Harchol-Balter, Downey
- 1996
(Show Context)
Citation Context ...b size distributions exhibit very high variability, and are best modeled by a Pareto distribution, a LogNormal, or a Weibull with high C 2 X . Some examples include Unix process CPU requirements (see =-=[1, 2]-=-) (where C2 values of 25 to 49 have been measured), sizes of files transferred through the Web [3, 4], sizes of files stored in Unix filesystems [5], durations of FTP transfers in the Internet [6], an... |

271 |
Queueing Systems, Volume II: Computer Applications
- Kleinrock
- 1976
(Show Context)
Citation Context ... jobs by the operating system CPU, whereby the CPU rotates among the jobs in the queue in a cyclic fashion, giving each job only a small fixed quantum of service before moving on to the next job, see =-=[9]-=-. If one imagines the limit as the quantum size goes to zero, one gets Processor-Sharing. Besides its use in modeling CPU processing, PS scheduling is also a good model of equal bandwidth sharing of a... |

216 |
Theory of Scheduling
- Conway, Maxwell, et al.
- 1967
(Show Context)
Citation Context ...n M/G/1 queue under the above scheduling policies. While we don’t have room to prove these results, we will provide high-level intuition behind each formula. We refer the reader to references such as =-=[12, 13, 14]-=- where complete proofs of these results can be found. While the formulas for response time that we present below are all well-known, the comparative ranking between the different scheduling policies i... |

155 | Analysis of SRPT scheduling: investigating unfairness
- Bansal, Harchol-Balter
- 2001
(Show Context)
Citation Context ...mbling block by showing that the “size” of a job, in this case the time to retrieve a file, is well-approximated by the size of the file, which is known. The authors deal with the starvation issue in =-=[26, 27]-=-, where they prove a counter-intuitive theorem, showing that, when job sizes follow a heavy-tailed distribution (like the Pareto or Bounded-Pareto), the expected response time of every job (including ... |

128 | Heavy-tailed probability distributions in the World Wide Web - Crovella, Taqqu, et al. - 1998 |

123 |
Load-balancing heuristics and process behavior
- Leland, Ott
- 1986
(Show Context)
Citation Context ...b size distributions exhibit very high variability, and are best modeled by a Pareto distribution, a LogNormal, or a Weibull with high C 2 X . Some examples include Unix process CPU requirements (see =-=[1, 2]-=-) (where C2 values of 25 to 49 have been measured), sizes of files transferred through the Web [3, 4], sizes of files stored in Unix filesystems [5], durations of FTP transfers in the Internet [6], an... |

119 |
A proof of the optimality of the shortest remaining processing time discipline
- Schrage
- 1968
(Show Context)
Citation Context ...a job arrives with shorter processing requirement than the job in service, then it preempts the job in service. SRPT is known to minimize the mean response time and the queue length distribution, see =-=[11]-=-, however its use is limited to situations where job size and remaining size are known. 3PSJF – Preemptive-Shortest-Job-First. This preemptive policy is very similar to SRPT, but favors that job with... |

119 | Sizebased scheduling to improve web performance
- Harchol-Balter, Schroeder, et al.
- 2003
(Show Context)
Citation Context ...ovements, at the cost of no additional resources, have not been lost on the computer systems community, which has incorporated smart scheduling in a wide variety of applications including web servers =-=[20]-=-, routers [21], wireless networks [22], operating systems [23], and many more. Below we elaborate on just two such examples: In a series of papers [20, 24, 25], Harchol-Balter et al. look at the probl... |

111 | Connection scheduling in web servers
- Crovella, Frangioso, et al.
- 1999
(Show Context)
Citation Context ...de variety of applications including web servers [20], routers [21], wireless networks [22], operating systems [23], and many more. Below we elaborate on just two such examples: In a series of papers =-=[20, 24, 25]-=-, Harchol-Balter et al. look at the problem of speeding up the mean response time in retrieving files at a web server. They find that the bottleneck resource for static “Get FILE” requests is the limi... |

109 | On choosing a task assignment policy for a distributed server system
- Harchol-Balter, Crovella, et al.
- 1999
(Show Context)
Citation Context ... has been shown to be far smaller than that in the open systems we have considered thus far [32]. By contrast, scheduling can have huge impact in server farm architectures (multi-server systems), see =-=[33]-=-, where scheduling decisions need to be made both at the central router level and at the individual host level. Lastly, fluctuations in load, e.g. alternations between very high and low load, can also... |

98 |
Queueing Analysis: A Foundation of Performance Evaluation. Vol. 2. Finite Systems
- Takagi
- 1993
(Show Context)
Citation Context ...eriod in an M/G/1 queue (for any work-conserving policy), and B(x) to denote the length of a busy period conditioned on the fact that the starting job is of size x. Then, it is well known (see, e.g., =-=[15]-=-) that the expected length of B(x) and B increase with load ρ in the following way: x E[B(x)] = 1 − ρ E[B] = E[X] 1 − ρ (2) (3) First-Come-First-Served (FCFS) A job’s response time, T , under FCFS can... |

90 | Classifying scheduling policies with respect to unfairness in an M/GI/1
- Wierman, Harchol-Balter
- 2003
(Show Context)
Citation Context ...mbling block by showing that the “size” of a job, in this case the time to retrieve a file, is well-approximated by the size of the file, which is known. The authors deal with the starvation issue in =-=[26, 27]-=-, where they prove a counter-intuitive theorem, showing that, when job sizes follow a heavy-tailed distribution (like the Pareto or Bounded-Pareto), the expected response time of every job (including ... |

80 |
A proof of the queuing formula
- Little
- 1961
(Show Context)
Citation Context ...ber of arrivals between each completion is also stochastically the same across policies, resulting in the same stochastic process for the number of jobs in the system. It then follows by Little’s Law =-=[17]-=- that the mean response time is also the same for all such policies. Hence: E[T] RANDOM = E[T] LCFS = E[T] FCFS = E[X] + λE [ X 2] 2(1 − ρ) Note that although the distribution of the number of jobs in... |

61 | servers under overload: How scheduling can help
- SCHROEDER, HARCHOL-BALTER
- 2006
(Show Context)
Citation Context ...de variety of applications including web servers [20], routers [21], wireless networks [22], operating systems [23], and many more. Below we elaborate on just two such examples: In a series of papers =-=[20, 24, 25]-=-, Harchol-Balter et al. look at the problem of speeding up the mean response time in retrieving files at a web server. They find that the bottleneck resource for static “Get FILE” requests is the limi... |

58 |
Processor-sharing queues: Some progress in analysis, Queueing Syst
- Yashkov
- 1987
(Show Context)
Citation Context ...ete independence of the job size variability. This property is best illustrated by the perfectly horizontal line in Figure 4, showing the insensitivity of mean response time to C 2 . The survey paper =-=[18]-=- presents more results on PS. Preemptive-Last-Come-First-Served (PLCFS) The PLCFS policy has surprisingly little to do with the LCFS policy, although they are both based on serving the last job to arr... |

52 |
Unix file size survey
- Irlam
- 1993
(Show Context)
Citation Context ...mples include Unix process CPU requirements (see [1, 2]) (where C2 values of 25 to 49 have been measured), sizes of files transferred through the Web [3, 4], sizes of files stored in Unix filesystems =-=[5]-=-, durations of FTP transfers in the Internet [6], and CPU requirements for supercomputing jobs [7]. In this article we will therefore focus a lot of attention on the effect of job size variability. De... |

45 |
Scheduling multiclass single server queueing systems to stochastically maximize the number of successful departures
- Righter, Shanthikumar
- 1989
(Show Context)
Citation Context ...quivalent to serving that job that has the lowest expected remaining service requirement. FB has been proven to minimize the mean response time and queue length distribution among blind policies, see =-=[10]-=-. We now move on to policies which make use of a job’s size, favoring shorter jobs in some way. Given that typically the great majority of jobs are “short,” favoring short jobs reduces overall mean re... |

37 | Evaluation of task assignment policies for supercomputing servers: The case for load unbalancing and fairness
- Schroeder, Harchol-Balter
- 2004
(Show Context)
Citation Context ...asured), sizes of files transferred through the Web [3, 4], sizes of files stored in Unix filesystems [5], durations of FTP transfers in the Internet [6], and CPU requirements for supercomputing jobs =-=[7]-=-. In this article we will therefore focus a lot of attention on the effect of job size variability. Decreasing failure rate (hereafter referred to as DFR) is also prevalent in computer science applica... |

30 | Nearly insensitive bounds on SMART scheduling
- Wierman, Harchol-Balter, et al.
- 2005
(Show Context)
Citation Context ...5]. Also, while we have studied individual policies in this paper, there’s a growing body of work that deals with broad classes of scheduling policies. In particular the SMART class was introduced in =-=[31]-=- and defined so as to encompass a wide range of policies which favor short jobs. Scheduling theory also exists for architectures far more general than the M/G/1 queue. Many computer systems follow a c... |

25 | Competitive online scheduling for server systems. Performance Evaluation Review, Special Issue on New Perspectives in Scheduling - Pruhs - 2007 |

16 | Fundamental characteristics of queues with uctuating load
- Gupta, Harchol-Balter, et al.
- 2006
(Show Context)
Citation Context ... central router level and at the individual host level. Lastly, fluctuations in load, e.g. alternations between very high and low load, can also greatly intensify the effect of scheduling, as seen in =-=[25, 34]-=-. Finally, while everything we have mentioned so far involves a stochastic setting, there is an entire field of online scheduling where the metric is worst-case response time and policies are ranked a... |

15 | The foreground-background queue: A survey
- Nuyens, Wierman
- 2008
(Show Context)
Citation Context ...ased job size variability, particularly when increased variability implies even stronger DFR behavior, as in the case of the Weibull job size distribution or the Pareto distribution. The survey paper =-=[19]-=- presents more results on FB. As a final point, we note that the fact that the curves for FCFS, PS, and FB all cross in Figure 4 at the point where C 2 = 1 is no accident. For exponentially distribute... |

15 |
Performance modeling of LAS based scheduling in packet switched networks
- Rai, Urvoy-Keller, et al.
- 2004
(Show Context)
Citation Context ...he cost of no additional resources, have not been lost on the computer systems community, which has incorporated smart scheduling in a wide variety of applications including web servers [20], routers =-=[21]-=-, wireless networks [22], operating systems [23], and many more. Below we elaborate on just two such examples: In a series of papers [20, 24, 25], Harchol-Balter et al. look at the problem of speeding... |

15 | Size matters: Size-based scheduling for mpeg-4 over wireless channels
- Mangharam, Demirhan, et al.
- 2004
(Show Context)
Citation Context ... resources, have not been lost on the computer systems community, which has incorporated smart scheduling in a wide variety of applications including web servers [20], routers [21], wireless networks =-=[22]-=-, operating systems [23], and many more. Below we elaborate on just two such examples: In a series of papers [20, 24, 25], Harchol-Balter et al. look at the problem of speeding up the mean response ti... |

14 | Scheduling for Today's Computer Systems: Bridging Theory and Practice
- Wierman
- 2007
(Show Context)
Citation Context ...n M/G/1 queue under the above scheduling policies. While we don’t have room to prove these results, we will provide high-level intuition behind each formula. We refer the reader to references such as =-=[12, 13, 14]-=- where complete proofs of these results can be found. While the formulas for response time that we present below are all well-known, the comparative ranking between the different scheduling policies i... |

11 |
Closed versus open system models: a cautionary tale
- Schroeder, Wierman, et al.
- 2006
(Show Context)
Citation Context ...nd a new job creation is only triggered by a job completion. The impact of scheduling in closed-loop systems has been shown to be far smaller than that in the open systems we have considered thus far =-=[32]-=-. By contrast, scheduling can have huge impact in server farm architectures (multi-server systems), see [33], where scheduling decisions need to be made both at the central router level and at the ind... |

10 | Pbs: a unified priority-based cpu scheduler
- Feng, Misra, et al.
- 2007
(Show Context)
Citation Context ...n lost on the computer systems community, which has incorporated smart scheduling in a wide variety of applications including web servers [20], routers [21], wireless networks [22], operating systems =-=[23]-=-, and many more. Below we elaborate on just two such examples: In a series of papers [20, 24, 25], Harchol-Balter et al. look at the problem of speeding up the mean response time in retrieving files a... |

8 |
Scheduling in practice
- Biersack, Schroeder, et al.
- 2007
(Show Context)
Citation Context ... the more CPU a UNIX job has used so far, the more CPU it is expected to use in the future [1], and the more bytes a flow has transmitted so far, the more bytes it is likely to transmit in the future =-=[8]-=-. 3 Definition of Common Scheduling Policies In this section, we define the most common scheduling policies for a single-server system. These can be subdivided along two axes, see Figure 2. Blind poli... |

2 |
Online Lecture Notes for Performance Modeling Class at CMU. http://www.cs.cmu.edu/harchol/Perfclass/class05.html
- Harchol-Balter
(Show Context)
Citation Context ...n M/G/1 queue under the above scheduling policies. While we don’t have room to prove these results, we will provide high-level intuition behind each formula. We refer the reader to references such as =-=[12, 13, 14]-=- where complete proofs of these results can be found. While the formulas for response time that we present below are all well-known, the comparative ranking between the different scheduling policies i... |

2 | Tails in scheduling. Performance Evaluation Review - Boxma, Zwart - 2007 |

2 |
Fairness and classifications. Performance Evaluation Review
- Wierman
- 2007
(Show Context)
Citation Context ...se time tails for FCFS and SJF mimic the tail of the excess of the job size distribution. Other performance metrics, like “fairness,” have also recently become very important in computer systems (see =-=[27, 29]-=- for an overview). There are many scheduling policies which we did not have room to mention. In particular there are many variants of PS that come up in networking (see the survey [30]). We also have ... |

2 |
Urtzi Ayesta, Sem Borst, Vishal Misra, and Rudesindo Nunez-Queija. Beyond processor sharing
- Aalto
- 2007
(Show Context)
Citation Context ...ystems (see [27, 29] for an overview). There are many scheduling policies which we did not have room to mention. In particular there are many variants of PS that come up in networking (see the survey =-=[30]-=-). We also have not had room to delve into multi-class queues, where each class i of jobs has its own arrival rate λi and its own job size distribution, Fi, and jobs in one class might have preemptive... |

1 | Scheduling in practice. Performance Evaluation Review - Biersack, Schroeder, et al. - 2007 |