Powerful processors and high-bandwidth communications which are available provide a very large amount of cycles at supercomputer speeds. This can serve emerging data-intensive applications which require extremely high data rates. Workstations equipped with such processors are often utilized for a fraction of the time and sit idle otherwise. It may become possible to sell and buy computational power over the network in the same way that electrical power is sold and bought by utilities over the power grid today. Potential buyers are users in need of more cycles than those provided by their own workstations. A request made by a potential buyer can be directed to any of the sellers and a best match of buyer and seller (trading) has to be made. For scalability, traders use an extended version of the adaptive partitioning algorithm which may be applicable whenever mutual interest is clearly defined, i.e. a seller is interested in finding a buyer and vice versa. It is shown that server or client selection for buying or selling computation power over the network using adaptive partitioning is simple and efficient.

Distributed applications spanning multiple nodes should he capable of providing fast response. Computation servers equipped with powerful processors and large memories can improve performance by better utilising system resources such as offloading overloaded nodes. A computation service may comprise all the nodes of the system (in case all are equipped with the required resources) or a small subset such as provided by a pool of computation servers (as in Amoeba). In both cases there can be a large number of service providers (compute servers). A server selection service is required to choose the most suitable server to provide the service. This paper is concerned with the design of adaptive computation server selection with scale recognised as a primary design and implementation factor. Adaptive system partitioning into domains is advocated as a key design principle for scalability. The model is demonstrated with compute-server selection using adaptive partitioning compared to the same service using random selection and probing.>

A method for qualitative and quantitative analysis of load sharing algorithms is presented, using a number of well known examples as illustration. Algorithm design choice are considered with respect to the main activities of information dissemination and allocation decision making. It is argued that nodes must be capable of making local decisions, and for this efficient state, dissemination techniques are necessary. Activities related to remote execution should be bounded and restricted to a small proportion of the activity of the system. The quantitative analysis provides both performance and efficiency measures, including consideration of the load and delay characteristics of the environment. To assess stability, which is also a precondition for scalability, the authors introduce and measure the load-sharing hit-ratio, the ratio of remote execution requests concluded successfully. Using their analysis method, they are able to suggest improvements to some published algorithms.>