Divisible Load Scheduling in Distributed Computing Environments

When attempting to achieve a time-consuming computation it is often advantageous to split the computation amongst multiple machines. Splitting a computation leads to the question of how to divide the load amongst the machines. If we assume that the load can be arbitrarily split and that there are no precedence relations amongst pieces of the load, we call the load divisible. One example of a divisible load is trying to find something in a large search space. Here, each machine can be given a portion of the search space and can operate independently of the other machines. If machine speed is all that is taken into account, then it is easy to optimally split a divisible load amongst a network of machines – simply solve a set of linear equations. However, if we also consider the network topology and the time it takes to send the load on the links of the network, then finding the optimal division of the computation becomes much more interesting. In this paper we attempt to investigate how such an optimal division of the computation can be achieved. In the next section we specify the problem formulation and discuss previous work. Then, we show that under our assumptions the problem of finding the optimal load distribution for a divisible load on an arbitrary network is NP-hard. Finally, we discuss some heuristics to find good load distributions.