IBM Research has a distinguished history in the theory and practice of performance analysis, modeling, and optimization. Fundamental contributions have been made in a number of important areas, including product-form queueing networks, optimal control and scheduling in queueing networks, stochastic ordering and majorization, rare event and parallel simulation, matrix-analytic analysis of stochastic models, polling systems, and performance modeling tools. These have played a critical role in understanding important problems in the design, development, and management of complex systems.

A key focus area is stochastic models and queueing theory. Our analysis of data from a wide range of systems demonstrates complex arrival and service patterns that include short-range dependencies, long-range dependencies, heavy-tail distributions, and nonstationary effects. While such behaviors often cannot be addressed with traditional methods, we have demonstrated that these complexities can degrade performance up to several orders of magnitude. We have also obtained analytical results for stochastic models and queueing networks with these complex characteristics, by using approximate methods and asymptotic analysis.

Another focus is the control and optimization of performance and other measures in stochastic queueing networks. Recent and emerging applications, e.g., those based on the Internet, have raised issues regarding differentiated service and quality of service based on non-traditional metrics, e.g., service-level-agreements based on distributions as opposed to averages. This together with the complex arrival and service processes found in practice make many of the optimal control and resource allocation problems extremely difficult. We have established provably optimal control policies for various queueing networks, in some cases based on asymptotic solutions.

These theoretical results have been applied in many different areas, including traffic generation and benchmarking, model validation, capacity planning, workload and performance forecasting, power-consumption models, generating and serving dynamic content, resource control and management, cooperative caching, dynamic offload, as well as network and server design.

Performance management seeks to increase the level of automation for delivering services to end users and managing the quality of such services. Our research activities include investigating techniques for forecasting service problems using time-series models, exploring the application of data mining techniques to problems in availability and performance management, and developing generic adaptive agents for automated tuning of complex systems that includes the application of control theory.

We have also developed performance tools. This includes a recently developed set of tools to automate the measurement and analysis of end-to-end web page retrieval, recommend changes that improve performance, and estimate the benefits of proposed changes. We are also developing tools that provide continuous monitoring of web page retrieval requests, as well as complex e-business transactions for service-level management. Tools that support the design, analysis, and simulation of hardware systems are another active area of research. We also continue to develop performance-modeling tools that are based on various methodologies - for example, matrix-analytic methods - and that are used to solve a wide range of general stochastic models arising in the design, development, management of complex systems.

Given the interdisciplinary nature of our research, there are considerable interactions with colleagues in many different areas, including algorithms and theory, communications and networking, computer architecture, e-commerce, knowledge discovery and data mining, mobile computing, operations research, operating systems, and web.

Please contact Paridhi Verma to obtain copies of the Computer Science Brochure