Grids and Worldwide Computing

Members of Rensselaer’s Center for Pervasive Computing and Networking are involved in research projects to overcome many of the challenges that stand in the way of worldwide computing. In this vision, local computers – and even various mobile devices such as cameras, credit cards, or computer-enabled watches, rings and eyeglasses – come together to form a virtual organization. Such an organization could use the shared computing resources to form a “grid,” a virtual supercomputer that tackles difficult scientific problems. Or a virtual organization could bring together all parties to a transaction such as a house purchase, including real estate agents, lawyers, banks, a title search firm, insurance companies, mortgage brokers and issuing agencies, inspectors, taxing jurisdictions, and the title records office to complete a transaction entirely on-line. Although grid computing made a lot of progress in the recent years through projects such as Globus, the focus in Rensselaer’s Center is on more dynamic and autonomous environments in which the task allocation, migration, and fault tolerance are automatically supported.

Middleware Software for Grids
 

High-performance computing once relied almost exclusively on supercomputers and massively parallel machines. The increasing power of PCs and workstations, however, has made it possible to use clusters of these machines for parallel computing. Even more complex is metacomputing, the effort to connect hundreds, thousands, or even millions of computers through the Web to tackle problems of universal interest. One difficulty is ensuring efficient and fast access to a huge and widely distributed database. Boleslaw Szymanski, professor of computer science and founding director of Rensselaer’s Center for Pervasive Computing and Networking, and his colleagues are developing high-level services and protocols that significantly boost the performance of grid computing. Their replication management protocols offer high data availability, low bandwidth consumption, increased fault tolerance, and improved scalability. Dr. Szymanski collaborates with the GLOBUS Group to integrate his tools into the GLOBUS system for grid computing. He also works with IBM-Almaden, using this group’s Smart-Grid system as a test bed for his research. In a project designed to show the power of metacomputing, Dr. Szymanski and others at Rensselaer linked hundreds of heterogeneous computers to study twin primes, pairs of prime numbers that differ by only two, such as 3 and 5 or 11 and 13. They counted the number of twin primes, computed the sum of their inverses, and found the maximum distance between twins and the locations of the gaps for a range of numbers two orders of magnitude larger than the previous record-breaking effort. Using idle time on the processors of those who had volunteered to participate, they completed the calculation in two years rather than the 100 years it would have taken on a single workstation.
 

Programming Models and Tools for Dynamic Grids
 

Carlos Varela, assistant professor of computer science at Rensselaer, is developing programming models, languages, and tools for worldwide computing. For many of the applications now envisioned, a system is needed that can evenly distribute workloads and redistribute them as conditions change or devices move, that allows nodes randomly to come on and off line, that helps the server reconfigure itself as it is accessed by a range of devices, and that can run without interruption as the basic topology of the network evolves. In collaboration with Dr. Gul Agha at the University of Illinois at Urbana-Champaign, Dr. Varela uses the actor model, which provides a unit of abstraction for concurrency and mobility. Dr. Varela and his colleagues have developed SALSA (Simple Actor Language, System, and Architecture), an actor programming language, to support the specialized needs of mobile computing. In a collaboration with IBM, Dr. Varela is developing a transactor model for reliable Web computing, a model specifically designed to support e-business by building reliability into a system that may contain unreliable nodes or links. In other work supported by IBM, Dr. Varela is using the Eclipse open source code as a tool to help visualize and graphically represent distributed computations. Drs. Varela, Joseph Flaherty, and Szymanski are also working with a group of physicists and biologists to use a worldwide computing approach for two scientific computation applications – the search for missing baryons (sub-atomic particles that have not been observed) and a study of the mechanisms by which bacteria evolve.
 

Scalable Routing in Unreliable Grids
 

Alhussein Abouzeid, assistant professor of electrical, computer, and systems engineering at Rensselaer, is investigating methods of routing messages on the unreliable networks formed in grid computing. In such cases, it is necessary to find nodes where computing resources are available, to allocate jobs to resources in ways that maximize performance, and to find new routes when nodes become unavailable. One must do all of this while expending a minimum amount of overhead, defined as computing time and power, and the system must be scalable to infinitely large systems without deteriorating quality. Dr. Abouzeid analyzes the changes of topology as a random process, and applies information-theoretic principles to quantify the minimum amount of overhead (bits/sec). He introduced the principle of “statistical traffic localization” that governs the selection of interacting peers in a dynamic grid. He has shown that grids that are designed according to this principle are infinitely scalable to arbitrary large numbers while keeping the overhead bounded.
 

Virtual Embedding for World Wide Computing
 

Bulent Yener, associate professor of computer science, suggests bridging the gap between supercomputing and grid computing by creating a virtual mesh and embedding it on top of the arbitrary mesh formed by the heterogeneous group of computers used in grid computing. The goal is to embed a virtual supercomputer architecture with minimum interference. In a dynamic wireless network, for example, routing information is obtained for the mobile terminals and mobile base stations, and this information is partitioned into a hierarchical, distributed database, based on the virtual topology. He has shown that such a system can cope with a high degree of mobility in the system and still successfully deliver packets to their destinations.
 

Concept-Based Software for Embedded Systems
 

David Musser, professor of computer science, is one of the originators of the concept-based approach that has led to widely used generic software libraries for data processing, scientific computation, and graph computation, including the Standard Template Library, the Matrix Template Library, and the Boost Graph Library. He is extending these methods to produce generic libraries and high-level optimization for embedded computing systems. In the concept-based approach, software artifacts are systematically classified according to their formal requirements. Concept refinement uncovers basic relationships and makes it possible to develop generic components that produce optimal program designs. Dr. Musser proposes using traditional concepts such as Algorithm, Iterator, Functor, Adaptor, and Allocator. In collaboration with Drs. Luk and Szymanski, he is also developing new concepts, such as distribution coordinators and mobility negotiators, that are needed for distributed real-time and embedded computing. His strategy is to begin with real problems being tackled in the Center for Pervasive Computing and Networking and work upward toward generalized concepts.
 

Parallel Processing Algorithms for Eye Surgery
 

Rensselaer researchers are using high-speed parallel processing algorithms to develop an improved system for laser surgery on the retina. Their system gives surgeons real-time information to better aim the laser. Surgeons now must look at a small section of the eye and coordinate it with another map of the whole eye, a task so difficult that there is a 50 percent error rate. Chris Carothers, assistant professor of computer science, and collaborators have developed algorithms that take advantage the 4 MB pages provided by an Intel processor to continuously in real-time map blood vessels, locate landmarks, and coordinate images, keeping track of the laser’s position despite the unpredictable, random movements of the human eye during surgery.