This paper reports on a project concerned with various aspects of stream processing based on distributed components. In earlier work, we have explored dynamic reconfiguration, distributed management and a Web services interface for remote deployment and control of cloud-based streams. Here, we look at aspects of remote interaction between a ProActive distributed active object system and the University of Chicago Nimbus cloud, showing that it is a useful environment in which to explore effects of distribution on latency, and commenting on problems of interaction through a firewall. We summarize some performance results of stream processing in the Nimbus cloud, especially our mechanisms for latency tolerance in remote interaction, with network measurements to show their effectiveness. The major contribution of this paper is to look at aspects of scheduling our stream components in a cloud based computation, focusing on criteria for initial deployment of the stream on cloud resources, outlining attributes of cloud resources that can sensibly be used in making scheduling decisions, and working towards a "cost model" suitable for use by economic scheduling heuristics. For this, we make use of grid-bus; we introduce our model of interaction with it, our preliminary implementation, and demonstration of the use of grid bus for deployment.
展开▼
