First published 2 September 2013
Cooperation, Congestion and Chaos in Concurrent Computation
Mizuki Oka, Takashi Ikegami, Alex Woodward, Yiqing Zhu, Kazuhiko Kato
We are interested in understanding how conflicts for common resources can be resolved when concurrently selfish agents are in place. To answer this question, we investigate a manycore machine that performs concurrent operations. Even with the selfish and non-cooperative nature of computational processes, they successfully organize a whole task. More specifically, we use the almost lock-free (ALF) architecture, which enables effective concurrent computation on a many-core machine. A unique point of the ALF is that it performs operations on shared resources simultaneously without excluding each other. We conducted data management experiments by varying the different number of cores on a single machine and investigating the characteristic dynamics of when the highest performance is observed. We found that the temporal dynamics of the number of operations changes from noisy to bursty pattern at the optimal point. In other words, the optimal computation is found at the edge of chaos. We argue that species or agents that interact concurrently with others show chaotic behavior in a congestion sate, and the cooperative state is established in the chaotic state.