Authors: Naoki Yonezawa, Issei Kino, Koichi Wada
Addresses: Department of Information Sciences, Faculty of Science, Kanagawa University, 2946 Tsuchiya, Hiratsuka, Kanagawa 259-1293, Japan. ' Department of Information Sciences, Faculty of Science, Kanagawa University, 2946 Tsuchiya, Hiratsuka, Kanagawa 259-1293, Japan. ' Department of Computer Science, University of Tsukuba, Tenoudai 1-1-1, Tsukuba, Ibaraki 305-8573, Japan
Abstract: Since barrier synchronisation is a simple means to guarantee the order of data producing and data consuming, it is often used in parallel programmes. However, barrier synchronisation causes the processors| idle time to increase. To reduce the overhead of barrier synchronisation, we have proposed an algorithm which eliminates barrier synchronisations and evaluated its validity experimentally. In this paper, we model the behaviour of parallel programmes and stochastically analyse our algorithm. Using the behavioural model, we evaluated the execution time before eliminating barrier synchronisations as well as after eliminating barrier synchronisations. As a result, we confirmed the observations, which we have found experimentally, that is: 1) the ratio of improvement increases as the number of processors increases: 2) the balance of the load affects the improvement which is gained from eliminating barriers.
Keywords: probabilistic analysis; time reduction; eliminating barrier synchronisation; parallel programming; exponential distribution; Erlang distribution; hyper-exponential distribution; variation coefficient; maximum random variables; summing random variables; Laplace-Stieltjes transform; LST.
International Journal of Communication Networks and Distributed Systems, 2011 Vol.6 No.4, pp.404 - 419
Published online: 26 Feb 2015 *Full-text access for editors Access for subscribers Purchase this article Comment on this article