JuxMem: An Adaptive Supportive Platform for Data Sharing on the Grid. Scalable Computing: Practice and Experience

Abstra t. We address the hallenge of managing large amounts of numeri al data within omputing grids onsisting of a federation of lusters. We laim that storing, a essing, updating and sharing su h data should be onsidered by appli ations as an external servi e. We propose a hierar hi al ar hite ture for this servi e, based on a peer-to-peer approa h. This ar hite ture is illustrated through a software platform alled JuxMem (for Juxtaposed Memory), whi h provides transparent a ess to mutable data, while enhan ing data persisten e in a dynami environment. Managing the volatility of storage resour es is spe ially empha-sized. As a proof of on ept, we des ribe a prototype implementation on top of the JXTA peer-to-peer framework, and we report on a preliminary experimental evaluation. Key words. data sharing, grid, peer-to-peer, hierar hi al ar hite ture, JXTA. 1. Introdu tion. A major ontribution of the grid omputing environments developed so far is to have de oupled omputation from deployment. Deployment is then onsidered as an external servi e provided by the underlying infrastru ture, outside the appli ation. This servi e is in harge of lo ating and intera ting with the physi al resour es, in order to e iently s hedule and map the omputation. In ontrast, as of today, no su h sophisti ated servi e exists regarding data management on the grid. Paradoxi ally enough, omplex infrastru tures are available for transparent omputation s heduling on distributed sites, whereas the user is still left to expli itly store and transfer the data needed by the omputation between these sites. At best, advan e