Society increasingly depends on large distributed systems, such as the Internet and Web-based service-oriented architectures deployed over the Internet. Such systems constantly evolve as new software components are injected to provide increased functionality, better performance and enhanced security. Unfortunately, designers lack effective methods to predict how new components might influence macroscopic behavior. Lacking effective methods, designers rely on engineering techniques, such as: analysis of critical algorithms at small scale and under limiting assumptions; factor-at-a-time simulations conducted at modest scale; and empirical measurements in small test beds. Such engineering techniques enable designers to characterize selected properties of new components but reveal little about likely dynamics at global scale. In this paper, we outline an approach that can be used to predict macroscopic dynamics when new components are deployed in a large distributed system. Our approach combines two main methods: scale reduction and multidimensional data analysis techniques. Combining these methods, we can search a wide parameter space to identify factors likely to drive global system response and we can predict the resulting macroscopic dynamics of key system behaviors. We demonstrate our approach in the context of the Internet, where researchers, motivated by a desire to increase user performance, have proposed new algorithms to replace the standard congestion control mechanism. Previously, the proposed algorithms were studied in three ways: using analytical models of single data flows, using empirical measurements in test beds where a few data flows compete for bandwidth, and using simulations at modest scale with a few sequentially varied parameters. In contrast, by applying our approach, we simulated configurations covering four-tier network topologies, spanning continental and global distances, comprising routers operating at state-of-the-art speeds and transporting more than 105 simultaneous data flows with varying traffic patterns and temporary spatiotemporal congestion. Our findings identify the main factors influencing macroscopic dynamics of Internet congestion control, and define the specific combination of factors that must hold for users to realize improved performance. We also uncover potential for one proposed algorithm to cause widespread performance degradation. Previous engineering studies of the proposed congestion control algorithms were unable to reveal such essential information.

This content is only available via PDF.
You do not currently have access to this content.