## Bounds On Contention Management Algorithms

### Cached

### Download Links

Citations: | 12 - 6 self |

### BibTeX

@MISC{Schneider_boundson,

author = {Johannes Schneider and Roger Wattenhofer},

title = {Bounds On Contention Management Algorithms},

year = {}

}

### OpenURL

### Abstract

We present two new algorithms for contention management in transactional memory, the deterministic algorithm CommitRounds and the randomized algorithm RandomizedRounds. Our randomized algorithm is efficient: in some notorious problem instances (e.g., dining philosophers) it is exponentially faster than prior work from a worst case perspective. Both algorithms are (i) local and (ii) starvation-free. Our algorithms are local because they do not use global synchronization data structures (e.g., a shared counter), hence they do not introduce additional resource conflicts which eventually might limit scalability. Our algorithms are starvation-free because each transaction is guaranteed to complete. Prior work sometimes features either (i) or (ii), but not both. To analyze our algorithms (from a worst case perspective) we introduce a new measure of complexity that depends on the number of actual conflicts only. In addition, we show that even a non-constant approximation of the length of an optimal (shortest) schedule of a set of transactions is NP-hard – even if all transactions are known in advance and do not alter their resource requirements. Furthermore, in case the needed resources of a transaction varies over time, such that for a transaction the number of conflicting transactions increases by a factor k, the competitive ratio of any contention manager is Ω(k) for k < √ m, where m denotes the number of cores. 1

### Citations

819 | Transactional Memory: Architectural Support for LockFree Data Structures
- Herlihy, Moss
- 1993
(Show Context)
Citation Context ...onflicts early, i.e. for every conflict a contention manager must have the possibility to abort any of the conflicting transactions. 2 Related Work Transactional memory was introduced in the nineties =-=[8, 16]-=-. In 2003 the FSTM system was proposed [6] and also the Dynamic STM (DSTM)[7] for dynamic data structures was described, which suggests the use of a contention manager as an independent module. After ... |

565 | Software Transactional Memory
- Shavit, Touitou
- 1995
(Show Context)
Citation Context ...onflicts early, i.e. for every conflict a contention manager must have the possibility to abort any of the conflicting transactions. 2 Related Work Transactional memory was introduced in the nineties =-=[8, 16]-=-. In 2003 the FSTM system was proposed [6] and also the Dynamic STM (DSTM)[7] for dynamic data structures was described, which suggests the use of a contention manager as an independent module. After ... |

436 | Language support for lightweight transactions
- Harris, Fraser
- 2003
(Show Context)
Citation Context ...ention manager must have the possibility to abort any of the conflicting transactions. 2 Related Work Transactional memory was introduced in the nineties [8, 16]. In 2003 the FSTM system was proposed =-=[6]-=- and also the Dynamic STM (DSTM)[7] for dynamic data structures was described, which suggests the use of a contention manager as an independent module. After these milestones, a lot of systems have be... |

381 | A simple parallel algorithm for the maximal independent set problem
- Luby
- 1986
(Show Context)
Citation Context ...max S is a bound on how long it takes at most until the next MIS can be scheduled. So, how to obtain a MIS without any knowledge about the conflict graph? The well-known distributed algorithm by Luby =-=[15]-=- computes a MIS with high probability (i.e., 1− 1 n ) in time O(tmax S ·log n). Unfortunately, it requires the degree of each transaction. Our Algorithm RandomizedRounds (see Section 5.5) works for dy... |

366 | Software transactional memory for dynamic-sized data structures
- Herlihy, Luchango, et al.
- 2003
(Show Context)
Citation Context ...ility to abort any of the conflicting transactions. 2 Related Work Transactional memory was introduced in the nineties [8, 16]. In 2003 the FSTM system was proposed [6] and also the Dynamic STM (DSTM)=-=[7]-=- for dynamic data structures was described, which suggests the use of a contention manager as an independent module. After these milestones, a lot of systems have been proposed. An overview of design ... |

204 |
Advanced contention management for dynamic software transactional memory
- Scherer, Scott
- 2005
(Show Context)
Citation Context ...ew can be found in [3]. Most proposed contention managers have been assessed by specific benchmarks only, and not analytically. A comparison of contention managers based on benchmarks can be found in =-=[12, 10]-=-. The experiments yield best performance for randomized algorithms, which all leave a (small) chance for arbitrary large completion time. Apart from that, the choice of the best contention managervar... |

148 | Hybrid transactional memory
- Damron, Fedorova, et al.
- 2006
(Show Context)
Citation Context ...nagervaries with the considered benchmark. Still, an algorithm called Polka [12] exhibits good overall performance for a variety of benchmarks and has been used successfully in various systems, e.g. =-=[2, 10]-=-. In [10] an algorithm called SizeMatters is introduced, which gives higher priority to the transaction that has modified more (shared) memory. In an enhanced version of this paper [14] we show that f... |

78 | Toward a theory of transactional contention managers
- Guerraoui, Herlihy, et al.
- 2005
(Show Context)
Citation Context ...f selfishness among programmers on the makespan is investigated for various contention managers from a game theoretic perspective. The first analysis of a contention manager named Greedy was given in =-=[5]-=-, using time stamps to decide in favor of older transactions. Variants of timestamping algorithms had been known previously (also in the field of STM [12]). However, [5] guaranteed that a transaction ... |

64 | On chromatic sums and distributed resource allocation - Bar-Noy, Bellare, et al. - 1998 |

59 | Improved inapproximability results for MaxClique, chromatic number and approximate graph coloring
- Khot
- 2001
(Show Context)
Citation Context ...on and for different graphs [14, 19, 21]. Our algorithm RandomizedRounds essentially computes a O(max{∆, log n}) coloring for a graph with maximum degree ∆. An optimal coloring is hard to approximate =-=[13]-=- and thus practically any work of distributed computing settles for coloring a graph with O(∆) colors.Additionally, in some cases O(∆) colors are required for coloring a graph, i.e. for graphs contai... |

52 | A log-star distributed maximal independent set algorithm for growth-bounded graphs
- Schneider, Wattenhofer
- 2008
(Show Context)
Citation Context ...igh probability.We relate the problem of contention management to coloring, where a large amount of distributed algorithms are available in different models of communication and for different graphs =-=[9, 13, 15]-=-. Our algorithm RandomizedRounds essentially computes a O(max{∆, log n}) coloring for a graph with maximum degree ∆. For further related work in respect to online scheduling, transactional memory syst... |

51 | Time-based Transactional Memory with Scalable Time Bases
- Riegel, Fetzer, et al.
- 2007
(Show Context)
Citation Context ...) in [1]. In contrast to our contribution, access to a global clock or logical counter is needed for every transaction which clearly limits the possible parallelism with a growing number of cores. In =-=[11]-=- a scalable replacement for a global clock was presented using synchronized clocks. Unfortunately, these days most systems come without multiple clocks. Additionally, there are problems due to the dri... |

43 |
MetaTM/TxLinux: Transactional memory for an operating system
- Ramadan, Rossbach, et al.
- 2007
(Show Context)
Citation Context ...ew can be found in [3]. Most proposed contention managers have been assessed by specific benchmarks only, and not analytically. A comparison of contention managers based on benchmarks can be found in =-=[12, 10]-=-. The experiments yield best performance for randomized algorithms, which all leave a (small) chance for arbitrary large completion time. Apart from that, the choice of the best contention managervar... |

26 | Understanding tradeoffs in software transactional memory
- DICE, SHAVIT
(Show Context)
Citation Context ...ests the use of a contention manager as an independent module. After these milestones, a lot of systems have been proposed. An overview of design issues from a practical point of view can be found in =-=[3]-=-. Most proposed contention managers have been assessed by specific benchmarks only, and not analytically. A comparison of contention managers based on benchmarks can be found in [12, 10]. The experime... |

24 |
Weak graph colorings: distributed algorithms and applications
- Kuhn
- 2009
(Show Context)
Citation Context ...igh probability.We relate the problem of contention management to coloring, where a large amount of distributed algorithms are available in different models of communication and for different graphs =-=[9, 13, 15]-=-. Our algorithm RandomizedRounds essentially computes a O(max{∆, log n}) coloring for a graph with maximum degree ∆. For further related work in respect to online scheduling, transactional memory syst... |

16 | Transactional contention management as a non-clairvoyant scheduling problem
- Attiya, Epstein, et al.
- 2006
(Show Context)
Citation Context ...er and by an optimal offline scheduler, knowing all transactions in advance) is O(s 2 ), where s is the number of (shared) resources of all transactions together. The analysis was improved to O(s) in =-=[1]-=-. In contrast to our contribution, access to a global clock or logical counter is needed for every transaction which clearly limits the possible parallelism with a growing number of cores. In [11] a s... |

10 | Coloring unstructured wireless multi-hop networks
- Schneider, Wattenhofer
- 2009
(Show Context)
Citation Context ...conflicts to be added to the initial conflict graph, the bound of Theorem 5 (and also the one of Theorem 4) can be improved to O(maxT ∈ST (max{dT , log n} · tmax)), with an analogous derivation as in =-=[15]-=-. As explained in [14] the schedule N + T corresponds then to a coloring using O(max{∆, log n}) colors. Let us consider an example to get a better understanding of the bounds. Assume we have n transac... |

8 |
Distributed Computing and the Multicore Revolution
- Herlihy, Luchangco
(Show Context)
Citation Context ...rom a practical point of view can be found in [5]. Since the field of STM is relatively young, it is not surprising that the theoretical foundations need to be developed, as motivated for instance in =-=[1, 10]-=-. Most proposed contention managers have been assessed by specific benchmarks only, and not analytically. A comparison of contention managers based on benchmarks can be found in [18, 16]. The experime... |

5 | Needed: foundations for transactional memory
- Attiya
(Show Context)
Citation Context ...rom a practical point of view can be found in [5]. Since the field of STM is relatively young, it is not surprising that the theoretical foundations need to be developed, as motivated for instance in =-=[1, 10]-=-. Most proposed contention managers have been assessed by specific benchmarks only, and not analytically. A comparison of contention managers based on benchmarks can be found in [18, 16]. The experime... |

4 | Scheduling with conflicts: online and offline algorithms
- Even, Halldorsson, et al.
- 2008
(Show Context)
Citation Context ...a clique of size ∆. Contention management is related to online scheduling. In contrast to contention management, scheduling algorithms are centralized and assume known conflicts. For illustration, in =-=[7]-=- the competitive ratios of scheduling algorithms are given for conflicting jobs. Their algorithms are non-distributed and on arrival of a new job J all conflicting jobs of J are known all at once, tak... |

2 | Good Programming in Transactional Memory: Game Theory Meets Multicore Architecture
- Eidenbenz, Wattenhofer
- 2009
(Show Context)
Citation Context ...has modified more (shared) memory. In an enhanced version of this paper [14] we show that from a worst-case perspective Polka and SizeMatters may perform exponentially worse than RandomizedRounds. In =-=[4]-=- the effects of selfishness among programmers on the makespan is investigated for various contention managers from a game theoretic perspective. The first analysis of a contention manager named Greedy... |