Computing with Distributed Resources

Abstract

The metaphor “Network is the Computer” has received much attention lately. It is easy/hard to claim such an equivalence since neither term is defined precisely. It is easy to establish an equivalence by ignoring several key aspects of the network, such as the costs of remote data access, failures of network nodes and communication links, and the security issues inherent in distributed computing. We may then view the network as a repository of data, typically stored in distributed objects, which resembles the primary (and secondary) storage of a traditional computer. The underlying instruction set for the network computer consists of method calls on these objects; the effect of a method call is to modify the state of the object (similar to a store instruction in a traditional computer) and/or return some value (similar to a load instruction). Now, consider executing a high-level instruction, such as x:= f(y,z), on the network computer. The data represented by x, y and z may be stored at different machines, as well as the function f to be applied to y and z. An implementation of this statement has (1) to determine the sites of the data y and z, and, possibly, choose among several sites if the data are replicated, (2) choose the site where the actual computation has to take place, and (3) communicate the result to all sites where x is to be stored. The implementation could be even more elaborate when x , y and z are matrices, for example, and f is the matrix multiplication operator. Then, y and z may be sent as streams (by rows or columns), the appropriate computations are carried out, possibly by multiple computers, and the result x sent as a stream of element values as soon as such a value is computed.

What should be the structure of a high-level program for a network computer? Should a user be given the illusion that all data are locally available, there is no latency in accessing data, computations are not interleaved with computations performed by other users, and that there is never any failure? This is clearly the ideal view. Unfortunately, such a view cannot be currently supported by the internet. A user may have to confront the possibility that certain pieces of data may be unavailable, because the corresponding site has failed. The user may have to realize that certain pieces of data could be modified by other parties during a computation.

In this paper, we develop a theory and a set of notations to represent the counterpart of a function on a network computer, i.e., how x:= f(y,z) should be specified and computed. Since non-determinism is inherent in a network model of computation, f need not be a function. We propose a programming model which includes non-determinism as a central concept; we call f a task. Task calls can be nested and tasks can be called recursively. Additionally, a task can explicitly include time-outs in its specification. Tasks capture the essence of what is currently known as web services.

The traditional theory of transaction processing can be used in implementing tasks; specifically, in computing x:= f(y,z): (1) the computation of f can be regarded as atomic, and (2) eventually, either f(y,z) is assigned to x, or there is no change in the state of any object.