## Centralized Versus Distributed Schedulers for Multiple Bag-of-Task Applications (2006)

### Cached

### Download Links

- [www.cecs.uci.edu]
- [graal.ens-lyon.fr]
- [www.cs.ucsd.edu]
- [cseweb.ucsd.edu]
- [www.labri.fr]
- CiteULike
- DBLP

### Other Repositories/Bibliography

Venue: | IN INTERNATIONAL PARALLEL AND DISTRIBUTED PROCESSING SYMPOSIUM IPDPS’2006. IEEE COMPUTER |

Citations: | 26 - 12 self |

### BibTeX

@INPROCEEDINGS{Beaumont06centralizedversus,

author = {Olivier Beaumont and Larry Carter and Jeanne Ferrante and Arnaud Legrand and Loris Marchal and Yves Robert},

title = {Centralized Versus Distributed Schedulers for Multiple Bag-of-Task Applications},

booktitle = {IN INTERNATIONAL PARALLEL AND DISTRIBUTED PROCESSING SYMPOSIUM IPDPS’2006. IEEE COMPUTER},

year = {2006},

publisher = {Society Press}

}

### Years of Citing Articles

### OpenURL

### Abstract

Multiple applications that execute concurrently on heterogeneous platforms compete for CPU and network resources. In this paper we consider the problem of scheduling applications to ensure fair and e#cient execution on a distributed network of processors. We limit our study to the case where communication is restricted to a tree embedded in the network, and the applications consist of a large number of independent tasks that originate at the tree's root. The tasks of a given application all have the same computation and communication requirements, but these requirements can vary for different applications. Each application is given a weight that quantifies its relative value. The goal of scheduling is to maximize throughput while executing tasks from each application in the same ratio as their weights. We can

### Citations

1794 |
Data Networks
- BERTSEKAS, R
- 1992
(Show Context)
Citation Context ...pecified priority weights, { we use the objective func(k) α tion Maximize mink w (k) } . This function, called fair throughput in the following, corresponds to the wellknown MAX-MIN fairness strategy =-=[8, 18]-=- among the different applications, with coefficients 1/w (k). We will consider both centralized and decentralized schedulers. For smaller platforms it may be realistic to assume a centralized schedule... |

635 |
Equilibrium points in N-person games
- Nash
- 1950
(Show Context)
Citation Context ...ton [1, 2] to our framework. Last, as we have seen with the PBC heuristic, noncooperative approaches where each application optimizes its own throughput lead to a particularly unfair Nash equilibrium =-=[19, 13]-=-. An other approach could be a cooperative approach where several decision makers (each of them being responsible for a given application) cooperate in making the decisions such that each of them will... |

269 | Bandwidth sharing: objectives and algorithms
- Massoulié, Roberts
- 1999
(Show Context)
Citation Context ...pecified priority weights, { we use the objective func(k) α tion Maximize mink w (k) } . This function, called fair throughput in the following, corresponds to the wellknown MAX-MIN fairness strategy =-=[8, 18]-=- among the different applications, with coefficients 1/w (k). We will consider both centralized and decentralized schedulers. For smaller platforms it may be realistic to assume a centralized schedule... |

170 | Improved Approximation Algorithms for the Multi-Commodity Flow Problem and Local Competitive Routing
- Awerbuch, Leighton
- 1994
(Show Context)
Citation Context ...ur framework is however really challenging as both communications and computations are involved. A promising approach would be to adapt the decentralized multi-commodity flow of Awerbuch and Leighton =-=[1, 2]-=- to our framework. Last, as we have seen with the PBC heuristic, noncooperative approaches where each application optimizes its own throughput lead to a particularly unfair Nash equilibrium [19, 13]. ... |

112 | Scheduling Distributed Applications: The SimGrid Simulation Framework
- Legrand, Marchal, et al.
- 2003
(Show Context)
Citation Context ...ks: give priority to high-bandwidth children, and assign them tasks of larger communication-to-computation ratios. We evaluate the decentralized heuristics through extensive simulations using SimGrid =-=[17]-=-, and use a centralized algorithm (guided by the linear program solution) as a reference basis. The rest of the paper is organized as follows. In Section 2, we state precisely the scheduling problem u... |

89 |
Open Infrastructure for Network Computing,” http://boinc.berkeley.edu
- “Berkeley
(Show Context)
Citation Context ...ation ratio for all of its tasks. This ratio proves to be an important parameter in the scheduling process. This scenario is somewhat similar to that addressed by existing systems. For instance BOINC =-=[11]-=- is a centralized scheduler that distributes tasks for participating applications, such as SETI@home, ClimatePrediction.NET, and Einstein@Home. The scheduling problem is to maintain a balanced executi... |

80 | Y.: Scheduling strategies for master-slave tasking on heterogeneous processor platforms
- Banino, Beaumont, et al.
- 2004
(Show Context)
Citation Context ...licts. Further, each node is free to schedule its sends arbitrarily within a time period. Thus, this schedule is substantially easier than if the processors were connected as an arbitrary graph (c.f. =-=[3]-=-). Anodeatdepthddoesn’t receive any tasks during the first d − 1 time periods, so will only enter “steady state mode” in time period d. Similarly, the root will eventually run out of tasks to send, so... |

74 | Bandwidth-Centric Allocation of Independent Tasks on Heterogeneous Platforms
- Beaumont, Carter, et al.
- 2002
(Show Context)
Citation Context ...imal performance should not be expected. 4.3. Coarse-Grain Bandwidth-Centric (CGBC ) This heuristic (CGBC ) builds upon our previous work for scheduling a single application on a tree shaped platform =-=[6, 3]-=-. In bandwidth-centric scheduling, each node only needs to know the bandwidth to each of its children. The node’s own worker thread is considered to be a child with infinite bandwidth. The scheduler t... |

72 | Efficient collective communication in distributed heterogeneous systems
- Bhat, Raghavendra, et al.
- 1999
(Show Context)
Citation Context ...is executing). There are several scenarios for the operation of the processors, which are discussed in Section A3 of the Appendix. In this paper, we concentrate on the full overlap, single-port model =-=[9, 10]-=-. In this model, a processor node can simultaneously receive data from one of its neighbors, perform some (independent) computation, and send data to one of its neighbors. At any given time, there are... |

65 | A simple local-control approximation algorithm for multicommodity flow
- Awerbuch, Leighton
- 1993
(Show Context)
Citation Context ...ur framework is however really challenging as both communications and computations are involved. A promising approach would be to adapt the decentralized multi-commodity flow of Awerbuch and Leighton =-=[1, 2]-=- to our framework. Last, as we have seen with the PBC heuristic, noncooperative approaches where each application optimizes its own throughput lead to a particularly unfair Nash equilibrium [19, 13]. ... |

49 | A proposal for a heterogeneous cluster ScaLAPACK (dense linear solvers
- Beaumont, Boudet, et al.
(Show Context)
Citation Context ...mber of tasks of each type it should assign to each of its children each time period. Thereafter, no further global communication is required. Each scheduler thread uses a 1D load-balancing mechanism =-=[4]-=- to select a requesting thread and an application type. The 1D load-balancing mechanism works as follows: if choice i should be made with frequency f(i), and has already been made g(i) times, then the... |

29 | Autonomous Protocols for Bandwidth-centric Scheduling of Independent-task Applications
- Kreaseck, Carter, et al.
- 2003
(Show Context)
Citation Context ... program (3). This period may be huge, requiring the nodes to have unreasonably large buffers to ensure uninterrupted steady-state behavior. The problem of buffer size has already been pointed out in =-=[12, 7]-=-, where it is shown that no finite amount of buffer space is sufficient for every tree. It is also known that finding the optimal throughput when buffer sizes are bounded is a strongly NP-hard problem... |

16 | Independent and divisible tasks scheduling on heterogeneous star-shaped platforms with limited memory
- Beaumont, Legrand, et al.
- 2005
(Show Context)
Citation Context ... program (3). This period may be huge, requiring the nodes to have unreasonably large buffers to ensure uninterrupted steady-state behavior. The problem of buffer size has already been pointed out in =-=[12, 7]-=-, where it is shown that no finite amount of buffer space is sufficient for every tree. It is also known that finding the optimal throughput when buffer sizes are bounded is a strongly NP-hard problem... |

13 | Load balancing in distributed systems: An approach using cooperative games - Grosu, Chronopoulos, et al. |

10 |
Introduction to the Theory of Games; Concepts, Methods, Applications
- Forgó, Szép, et al.
- 1999
(Show Context)
Citation Context ...ton [1, 2] to our framework. Last, as we have seen with the PBC heuristic, noncooperative approaches where each application optimizes its own throughput lead to a particularly unfair Nash equilibrium =-=[19, 13]-=-. An other approach could be a cooperative approach where several decision makers (each of them being responsible for a given application) cooperate in making the decisions such that each of them will... |

8 | A strategyproof mechanism for scheduling divisible loads in distributed systems - Grosu, Carroll - 2005 |

3 | Scheduling multiple bags of tasks on heterogeneous master-worker platforms: centralized versus distributed solutions,” LIP
- Beaumont, Carter, et al.
- 2005
(Show Context)
Citation Context ...g a given processor, one sent and the other received. 1 Due to lack of space, we do not provide related work in this article but a thorough survey can be found in the extended version of this article =-=[5]-=-.2.2. Application Model We consider K applications, Ak, 1� k � K. The root node P0 initially holds all the input data necessary for each application Ak. Each application has a priority weight w (k) a... |

2 |
The Art of Computer Systems Performance Analysis : Techniques for Experimental Design, Measurement, Simulation, and Modeling
- Jay
- 1991
(Show Context)
Citation Context ...t of the tree while applications with a high communication-to-computation are mainly performed on the leftmost part, which is definitely not optimal on particular instances. 3 It is a well-known fact =-=[16]-=- that arithmetic average of ratios can lead to contradictory conclusions when changing the reference point. Therefore, we use a geometric average of ratios which is known to be closer to the general i... |