中国科学院软件研究所机构知识库

   Data distribution is NP-Complete problem which studies how to place data into resource nodes. Data distribution is one of key technologies of multi-nodes distributed systems, such as cloud computing, pervasive computing, grid computing, distributed
computing, and P2P network. Data distribution significantly influences system performance, reliability, trust, and resource allocation. With the development of cloud computing, cloud data distribution becomes an indispensable part of cloud computing and plays an important role in numerous areas, such as load balance, energy conservation, and system security.

   Current existing data distribution policies have gotten excellent result in whole time domain. They also have the similar performance in geographical distributed system.
Compared with the former geographical distributed systems, however, most of existing cloud infrastructures are large-scale clusters, of which bandwidth has been tremendously improved. As a result, data distribution policy needs to weaken the importance of some main factors which are influenced by bandwidth. Additionally, data distribution policy which is in balance in the whole time domain may has an unbalanced load in each shorter time domain. And bottleneck of system may derive from such unbalanced load.

   After studying kinds of data distribution policies, according with the loose characteristic of cloud data, pattern of accessing data in short time domain, and infrastructure of cloud, mathematical model of data distribution process is proposed.
Cloud data distribution system, centralized and distributed data distribution policies which are based on time sequence evaluation are designed and implemented. These policies split the whole time domain into time sequences, transform multi-objective optimization problem to single-objective problem, and use feedback of evaluating result to adjust datasets of nodes. These policies balance load of every node in each short time domain. According to lots of random simulation and cloud environment
experimentation, data distribution policy based on time sequence evaluation has gotten better result than other common data distribution policies in three kinds of evaluations: system balanced in entire time domain, system balanced in each short time domain, and the maximum wave peak of accessing resource nodes in entire time domain. Compared with centralized policy, distributed policy is significantly improved on evaluation of adjusting data size.