Abstract
XML-based publish/subscribe (pub/sub) systems have been receiving a great deal of attention from the academic community and the industry. This research focuses on efficient pub/sub systems and considers the devising of XML-based pub/sub systems from the perspective of subscription/query and publication message delivery. Existing research mainly focuses on the efficiency of XML publication message filtering algorithms. Not much research, however, has considered using the system or the communication model in the context of XML publication messages delivery. This paper presents innovative XML delivery techniques, the cross-layer model and the peer model; both techniques make use of publisher and customer edge brokers for efficient XML subscription aggregation and publication message delivery. The primary contribution of the proposed models is the reduction of the number of XML publication message filtering and XPath query aggregation operations performed in the conventional filter-based XML multicast model, which has a high computational overhead. The main idea is to store user subscriptions at customer and publisher edge brokers which are either directly connected or close to the subscribers and the publisher, respectively. We have performed a number of experiments within a controlled local area network (LAN) environment for demonstration of the basic concepts underlying the techniques and in the Amazon cloud environment that emulates the wide area network (WAN). Both the cross-layer and the peer models can reduce the end-to-end (E2E) delay in message delivery. For example, the results obtained from experiments in a LAN demonstrate several-fold performance improvement in E2E delay for both the cross-layer and the peer models compared to the conventional filter-based XML multicast model, and the results using the cloud show an improvement as high as 64% in E2E delay for the peer model over the multicast model.
6. Conclusions and future directions
XML pub/sub systems are important applications and the performance of such systems is important as the data volume keeps increasing. Performance improvement can be realized with various techniques. This paper focused on performance improvement from the communication models perspective. Effective communication models have potential to significantly improve the performance, as stated in Section 1. This paper presented two communication models for XML pub/sub services, i.e., the cross-layer model and the peer model. We then compared these two models with the conventional filter-based XML multicast model. The conventional XML multicast model builds an application-layer broker overlay above the network layer. The approach forwards the minimal number of XML messages to subscribers at the expense of two high computational cost of XML query aggregation and message filtering operations performed repeatedly at each hop from a PE broker to a CE broker. Both the proposed cross-layer model and the peer model limit the two expensive XML broker operations to be performed only at PE and CE nodes. However, the cross-layer model still relies on XML-capable brokers for message forwarding. On the other hand, the peer model does not establish and maintain a multicast tree, and message forwarding between PE and CEs is realized using the regular network routing technologies rather than following the multicast tree paths. The peer model offers several advantages: low computational cost (both query aggregation and publication filtering operations), easy deployment and management without the need of setting up application-layer XML-capable brokers, high network efficiency, support for frequent user subscription changes, and resilient of broker failures (assuming no failures occur at PE and CEs). An XML pub/sub prototype was devised and the different routing models were evaluated in a LAN environment and on the Amazon cloud centers around the world to emulate a WAN environment where propagation delay becomes the dominant factor. We compared three different models through extensive experimentation. The results showed that the cross-layer model and the peer model significantly outperformed the multicast model in a LAN environment primarily due to the elimination of the expensive filtering operation performed on each broker. Further, the peer model has a notably lower E2E delay than the other two models in a WAN environment as a result of eliminating the filtering operation performed at intermediate brokers and shorter paths used at the network layer than that of the application layer. The focus of this paper was on publication message forwarding. Integrating the publication message routing and user query aggregation techniques into a prototype and evaluating its performance with real workload warrants investigation.
