Atsushi Kubota

Identifying propagated response delays in performance monitoring of n-tier applications

One of the significant challenges on performance monitoring of an n-tier system is the “response delay propagation”, in which a response delay in a component server is propagated to other component servers due to the invoking relations among request types in different component servers of the system. It leads the operations manager of the system to misdiagnose the location of source delays and results in wasting time to investigate the root cause. We developed a response delay monitoring system that helps the operations managers distinguish the source delays from many other propagated delays. The system is able to build a model of invoking relations among request types in different component servers and use the model to diagnose the response delay propagation and pin-point the location of source response delays. To obtain such invoking relations among request-types from black-box component servers in an n-tier system, we propose a novel invoking relation estimation method which can achieve high accuracy of true invoking relation among request types by eliminating the negative influence of two spurious correlation factors through partial correlation analysis. We implemented the response delay monitoring system and evaluated the effectiveness of our invoking relation estimation method on a real in-company n-tier system which has thousands of request-types in each tier. The result (over 90% in precision) confirms our estimation method can effectively capture invoking relations in an n-tier system.

Grid Service Platform: Design and Implementation of Grid Middleware for Telecom Carriers

We have developed the Grid Service Platform (GSP), which is a grid middleware for telecom carriers. GSP can support not only non-interactive batch style services, but also interactive real-time services. Moreover, GSP attains autonomous resource sharing between services based on the priority of each service. We conducted a field trial of GSP on a testbed that consisted of three sites in Japan and France. Two different types of services were implemented on GSP: a video conferencing service (interactive) and a batch queuing service (non-interactive). As a result, the performance of both services were simultaneously enhanced. The session capacity of the video conferencing service was improved by up to 30%, while the total execution time of batch jobs was reduced by 12%.