Effective use of the recent innovations in computer architecture can be limited by difficulties in writing functionally correct parallel applications that also achieve high performance. Hybrid algorithms combine the advantages of more than one computing method to obtain the desired results.

The AHPCRC team at UTEP has developed a hybrid method that enables a user to choose from among several global stochastic techniques to search the parameter space for possible minima. The user has the option to use a parallel multi-start (i.e., more than one initial guess). Most of the global methods used here do not rely on derivatives, and thus do not need to know the direction in which a function is increasing or decreasing in order to work. The global search produces a set of target regions where the global optimal solution may lie. Data points from these regions are filtered to produce a surrogate model, which behaves in a mathematically similar fashion to the function of interest, while demanding less in terms of computational resources. The surrogate model is used to perform local searches, using an algorithm developed by the UTEP team. This method calculates gradients and evaluates whether a given gradient is steep enough to lead to a global minimum. The algorithm may also include physical bounds to ensure that the resulting approximate solutions are realistic.

The UTEP AHPCRC team is working to make it easier to implement HPC applications on highly parallel systems. They are developing and demonstrating a practical migration path from current programming approaches to a transaction-based model. Introducing transactions (individual small operations) as the key abstraction for expressing parallelism facilitates maintaining a computer system in a known, consistent state by ensuring that interdependent operations are either all completed successfully or all canceled successfully. The team is developing a simple distributed-memory programming model that can scale to systems with thousands of processors.

One area of particular interest is estimation theory—a branch of statistics that is often used to assist in interpreting the results of scientific experiments. Mathematical calculations use observable information as input to produce an approximation to the parameter of interest when an exact solution is not possible. Such techniques are especially useful for signal processing and telecommunications problems, as well as for scientific applications with irregular and adaptive behavior.

The hybrid optimization methods developed as a result of this project are being applied to Stanford's AERO-F computational fluid dynamics code, which is used in AHPCRC Technical Area 1 to model flapping and twisting wings for micro-aerial vehicles. Hybrid optimization is also used for hydraulics modeling applications of interest to the U.S. Army Corps of Engineers.

Read more about this project...