Gossip-based Strategies in Global Optimization

DIT-PRJ-09-013

Status NOT active project
DISI role Partner
Project type Research Project
Dimension International
Acquisition date 2007-12-10
Start date 2008-02-15
End date 2008-09-12
SAP code 30101860

Project details

Project astract In all areas of engineering, including space activities, finding the global optimum of various numerical and combinatorial problems is a central task. This task can be extremely time-consuming and compute-intensive, which requires the optimal utilization of available resources. Among such resources, desktop GRID-like environments in large organizations, such as ESA, represent a powerful yet low cost asset. In such systems a large number of heterogeneous and unreliable desktop computers are organized into a network with the help of basic resource management services. Unlike shared memory architectures or dedicated clusters with high speed and reliable interconnections, desktop grid systems might have relatively slow and unreliable links, and the availability of the resources at a specific computer might vary significantly with time in an unpredictable way due to uncontrollable factors external to the system. In such a system, novel approaches to the parallelization of global optimizers are necessary. We propose to tackle this problem during the project via considering various global optimization methods, and proposing parallel versions with the help of control mechanisms and techniques known from P2P computing and self-organizing distributed systems, such as gossip, diffusion, aggregation, and so on. Our goal will be to develop adaptive methods that can achieve speedups similar to that of a fully reliable shared memory system or a cluster, despite all the unreliable components and the potentially large scale of the desktop grid system.
Fundings 15000 €
Partners
  • DIT - UniTN
  • Hungarian Acad. Sci. and University of Szeged

DISI Sub-project details

Project astract In all areas of engineering, including space activities, finding the global optimum of various numerical and combinatorial problems is a central task. This task can be extremely time-consuming and compute-intensive, which requires the optimal utilization of available resources. Among such resources, desktop GRID-like environments in large organizations, such as ESA, represent a powerful yet low cost asset. In such systems a large number of heterogeneous and unreliable desktop computers are organized into a network with the help of basic resource management services. Unlike shared memory architectures or dedicated clusters with high speed and reliable interconnections, desktop grid systems might have relatively slow and unreliable links, and the availability of the resources at a specific computer might vary significantly with time in an unpredictable way due to uncontrollable factors external to the system. In such a system, novel approaches to the parallelization of global optimizers are necessary. We propose to tackle this problem during the project via considering various global optimization methods, and proposing parallel versions with the help of control mechanisms and techniques known from P2P computing and self-organizing distributed systems, such as gossip, diffusion, aggregation, and so on. Our goal will be to develop adaptive methods that can achieve speedups similar to that of a fully reliable shared memory system or a cluster, despite all the unreliable components and the potentially large scale of the desktop grid system.
Fundings 15000 €
Manager Alberto Montresor