Masha Sosonkina, Principal Investigator

Providing Dynamic Network Information to Distributed Applications

Clusters of personal computers provide an inexpensive alternative to high performance computing. However, the sustained performance of these computational and network resources is difficult to ensure for the total execution time of an application. Due to numerous factors, sudden changes in the amount of available resources are common, and this situation affects the running application.

To cope with fluctuations in hardware performance, an application may need to have adaptive capabilities. This is especially vital for scientific applications, since the majority of them are characterized by long execution times and by expensive recoveries from failure.

These researchers investigated the interaction between running scientific applications and the network interconnecting distributed computing resources. In particular, these researchers applied iterative methods to solve linear systems arising in static tire equilibrium computation. The heterogeneous material properties, nonlinear constraints, and a threedimensional finite element formulation made the linear systems arising in tire design difficult to solve by iterative methods. An analysis of the matrix characteristics helped understand this behavior. Focusing on two preconditioning techniques, a variation of an incomplete factorization with threshold and a multilevel recursive solver, the researchers adapted these techniques in a number of ways to work for a class of realistic applications.

It was found that these preconditioners improve convergence only when a rather large shift value is added to the matrix diagonal. A combination of other techniques—such as filtering of small entries, pivoting in preconditioning, and a special way of defining levels for the multilevel recursive solver—were shown to make these preconditioning strategies efficient for problems in tire design. Comparisons were made between these techniques, and their applicability was assessed in circumstances where the linear system difficulty varied for the same class of problems.

In related research, this group investigated means to improve the performance of distributed- memory architectures used by scientific applications. Distributed computations are widely used to solve large-scale realistic problems in scientific applications. By using standard communication libraries, such as Message Passing Interface (MPI), distributed scientific applications are made portable across various network and shared-memory interconnections of processors. For distributed-memory architectures, the performance of applications may vary greatly due to the different performances of various network technologies. Not only static (configuration) network characteristics, but also dynamically changing conditions such as network load, may affect the performance. For an application, one way to react to the dynamic network changes is to adjust its own communication needs. However, distributed (scientific) applications typically have no embedded means to learn the current network conditions. Therefore, it is desirable to have a mechanism that provides the network information transparently to the application programmer or user, so that the burden of handling the low-level network information is shifted to a network developer.

Research Group

Devdatta Kilkarni, Graduate Student Researcher