Intel Research Seminar


DATE: August 25, 2003
TIME: Noon - 1:30 pm
PLACE: Intel Seminar (417 S. Craig Street - 3rd Floor)
INTEL EVENTS PAGE: http://www.intel-research.net/pittsburgh/events.htm

SPEAKER:
John Reumann
University of Michigan

TITLE:
Improving System Support for the Management of Multi-Tiered Services

ABSTRACT:
Modern Internet-services are often multi-tiered. This means that front-end services (e.g. HTTP) act as gateways to a number of (potentially also multi-tiered) back-end services (e.g. databases). Such systems are notoriously difficult to manage because activities triggered by the front-end hosts propagate across multiple hosts and loosely-integrated services. This talk will introduce a generic system-layer framework, Stateful Distributed Interposition (SDI), for the policing of multi-tiered systems because dealing with tens or hundreds of application-specific management settings is impractical.

SDI identifies the following six interdependent mechanisms as fundamental building blocks for the design of system management features in multi-tiered server systems: interposition, system object tagging, rule-based classification, tag management, tracking, and tag-sensitive policy enforcement.

The later part of this talk highlights how tagging, policy enforcement, and tracing can be used to achieve enforcable resource partitions among activities that share some back-end services. This potentially reduces software licensing and administration costs (service sharing), facilitates pay-per-use type outsourcing models (activity-to-partition mapping), and reduces the need for excessive over-capacity (partition
resource insulation).

Since the above configuration of resource partitions in a multi-tiered system requires some understanding of the interactions among component services, the Performance Map tool is introduced. This tool generates maps of the interactions among services and system components and allows assessing their response time contribution to specific service classes. Besides being useful in the configuration of our own distributed resource partitioning scheme, this tool aids bottleneck detection and root-cause performance failure diagnostics without requiring the source codes of applications. This may greatly improve the productivity of system administrators.

BIO:
John Reumann is member of the University of Michigan'' Real-Time Computing lab. His research interests include operating systems and middleware software, distributed systems, security mechanisms, performance management, monitoring, computer networks, and online configuration adaptation. At the University of Michigan, John completed his PhD and MS in Computer Science in 8/2003 and 4/1999 respectively. He completed the equivalent of a B.S. in Computer Science at the University of Bonn, Germany, in 1997. In 2002 he founded NQ Technologies, seeking to commercialize his thesis research by delivering basic system management infrastructure software into the system management tools and services industries.

For Further Seminar Info:
Contact Kim Kaan, 412-605-1203, or visit http://www.intel-research.net.

SDI Home: http://www.pdl.cmu.edu/SDI/