Developing distributed applications for the Grid is a difficult task, since several concerns intertwine within them and hence must be coped with simultaneously. Moreover, when deploying application classes, several needs should be considered, such as their characteristics, user preferences, and runtime conditions of available hosts. RexMidas is a framework that automatically transforms a centralised Java application into a distributed, possibly Grid-enabled, version. RexMidas' support for remote allocation of, and communication among, class instances, is transparent, in that original application classes need not be modified (or available) at the source level. For remote allocation RexMidas adopts policies that try to strike a balance between choosing capable hosts, equipped with the necessary libraries, and minimising the overhead of communication between remote objects.

In a typical grid environment, replicas are spread over storage hosts, so as to increase availability and reduce files transfer times. Generally, the grid middleware lets users choose the storage host that will be serving the replica. However, having users able to choose could adversely affect network conditions and increase transfer times. This paper proposes two replica selection policies that choose one among the storage hosts holding a requested replica. The selection policies try to minimise file transfer time by extracting the least loaded host or the closest among unloaded hosts. It is shown that overall performances for transferring replicas depend on network topology, latency and bandwidth of links, and load of storage hosts and links.

This paper presents some experimental results on the evaluation of a resource finding algorithm for a peer-to- peer Grid. The algorithm, which is based on "spatial com- puting" concepts, models a Grid as a P2P overlay network, on which resources are obtained by a suitable navigation strategy that forwards a request from node to node, un- til one is found offering the desired amount of resources. The algorithm is experimentally evaluated by means of a software simulator, and fully characterised by assessing the performances afforded by different navigation strategies and the influence of several parameters, such as the number of nodes, their workload, the origin (node) of the resource request, etc.