Vadim E. Kotov

Systems of systems as communicating structures

By Systems of Systems (SoS) we mean large-scale distributed systems the components of which are complex systems themselves (e.g. enterprise intranets). Communicating Structures are represent SoS in a uniform, systematic way as composition of a small number of basic system objects and notions.

Fibre channel fabrics: evaluation and design

We consider how fibre channel switches can be cascaded to form a fibre channel fabric. We begin with an analytical model of topology performance that provides a theoretical upper bound on fabric performance and a method for the practical evaluation of fabric topologies. We then consider the prevention of buffer-cycle deadlock in fibre channel networks. We show that though some topologies implicitly do not deadlock, these topologies have inferior throughput, and we present a technique for optimizing performance that takes into account both deadlock and traffic balancing.

A framework for analyzing and organizing complex systems

The paper discusses a framework and technologies enabling the quantitative analysis, organization and optimization of large-scale, globally distributed enterprise and e-services systems. The goal is to organize complex systems in such ways that traffic can be better explained, predicted and controlled at application and service layers rather than at the network layers. Our work approaches higher system perspectives where architectural decisions are made about the overall organization of work and task flows, the global placement of data and applications, etc. Those decisions are significant for the traffic induced in the system later on. Little support is provided today for designing and evaluating large-scale systems from these perspectives, primarily caused by the difficulties in developing realistic computerized models reflecting the dynamic behavior of services and applications. We reduce complex environments to uniform representations of resource demands and capacities and use them to improve the overall system organization. Case studies with earlier versions of our approach have been carried out with two corporate partners, which are discussed at the end.

Resource-sharing and service deployment in virtual data centers

The expectation of a global presence of services leads to the need for large numbers of service instances allocated in a multitude of regional data centers in order to provide sufficient service capacity close to where the demand occurs. Scale of service instances is anticipated growing /spl Gt/10/sup 4/ raising new challenges for control and management. Pragmatically, it must become much easier to deploy service instances in data center, allocating resources, sharing them, installing and configuring data and software needed for service instances and integrating them into a singular service that appears to a consumer. Adjusting numbers and locations of service instances is seen as a basic control mechanism in order to follow regional or temporal fluctuations in demands. The paper proposes a new concept of virtualizing whole data center environments and quickly deploying massive amounts of service instances. A virtualization layer takes care of resource allocation from different data center locations and all specifics when service instances are allocated in a particular data center. Virtualized data centers provide a consistent operating environment spanning multiple physical data center locations for the whole family of service instances. And vice versa, physical data centers host several execution environments for different services. After a brief discussion of challenges coming with the scale of service instances we anticipate, the paper overviews virtual data centers and discusses one aspect in more detail how massive amounts of service instances can be deployed using a recursive approach.

Modeling a scalable high-speed interconnect with stochastic Petri nets

This paper presents an approach to using Stochastic Petri nets to model large-scale concurrent systems, in our case, a scalable computer interconnect. We show how Stochastic Petri net models can exploit the symmetry of the system to construct a tractable, but approximate, analytic model, and that they can yield results very close to those of a detailed simulation model, with much less computational effort.

Service-centric globally distributed computing

An automated service demand-supply control system can improve a large-scale grid infrastructure comprising a federation of distributed utility data centers.

Optimization of E-service solutions with the Systems of Servers library

The paper describes a modeling library for analysis and optimization of distributed service systems: enterprise computing infrastructures, E-commerce, and E-service systems. The library uses a small number of basic notions, such as service, server, cluster (of servers), clients and message, which allow describing, analyzing and optimizing various system configurations and distributions of services among servers. The service requests and responses are modeled by messages traveling in the system and using its common resources. Clients send their requests to servers, which either return responses or may issue secondary requests for other services. To find the optimal deployment of services among servers, different partition policies can be deployed including partitions based on genetic algorithms. The library is built on top of Communicating Structures Library (CSL), a modeling environment for large-scale distributed systems (V.E. Kotov, 1998). In its turn, Systems of Servers is a base for specialized libraries that target specific classes of service systems.

On Net Modeling of Industrial Size Concurrent Systems

The paper discusses the modeling of OLTP (On-Line Transaction Processing) using colored Petri nets and the Design/CPN tool. We have performed industrial-sized simulation of tens of thousands of transactions running on databases with millions of records. OLTP applications were chosen because the OLTP workloads emphasize update-intensive database services with up to thousands of concurrent transactions per second. The goal of the experiment was to develop a methodology for the net modeling of OLTP-like applications with respect to the ability of different computer systems to meet the OLTP benchmarks requirements. The effort spent on the successful modeling of such large systems resulted in some methodological conclusions which required us to avoid straightforward approaches to the net modeling and to develop a combined technique of using nets together with embedded programming for simulation and analytical modeling.

Self-Organizing Control in Plantetary-Scale Computing

The explosion in globally connected devices, computers, and services ultimately evolves into very large-scale, inter-connected systems approaching a new era of planetary-scale computing. The backbone of planetary-scale computing is a global network of data centers. The paper provides an overview of research at HP Laboratories exploring new approaches under such a scenario based on HPs virtual data center platform enabling large-scale distributed virtual data center environments. Our approach is founded in a service-centric system view, and we explore agent technology for resource control, system organization, and balancing resource demand and supply.

Adaptive service placement algorithms for autonomous service networks

Motivated by trends in the industry towards transforming IT in large integrated service networks, this paper describes algorithms for the adaptive placement of “services” (as abstractions of collections of applications) in networks of “servers” (as abstractions for locations where services can be hosted). Networks comprised of interacting services as the foundation is also a vision pronounced by the Grid [9]. Manageability and “self-operation” of Grids is highly desirable. We analyze the requirements for algorithms one specific problem: the service placement problem. We discuss algorithms that neither require central control nor complete information about the system state. Algorithms are performed on a distributed overlay structure which summarizes load conditions in the underlying service network. The presented algorithms fulfill tasks of making initial placement decisions as well as initiating rearrangements when imbalance is detected. Presented algorithms have different characteristics regarding the tradeoff between accuracy (or quality) of a placement decision and its timeliness within which a decision can be made determining responsiveness.