Automatic peptide identification from collision-induced dissociation tandem mass spectrometry data using optimization techniques is made difficult by large plateaus in the fitness landscapes of scoring functions, by the fuzzy nature of constraints from noisy data and by the existence of diverse but equally justifiable probabilistic models of peak matching. Here, two different scoring functions are combined into a parallel multi-objective optimization framework. It is shown how multi-objective optimization can be used to empirically test for independence between distinct scoring functions. The loss of selection pressure during the evolution of a population of putative peptide sequences by a Pareto-driven genetic algorithm is addressed by alternating between two definitions of fitness according to a numerical threshold. Copyright © 2005 John Wiley & Sons, Ltd.

Summary form only given. Automatic de novo peptide identification from collision-induced dissociation tandem mass spectrometry data is made difficult by large plateaus in the fitness landscapes of scoring functions and the fuzzy nature of the constraints that is due to noise in the data. A framework is presented for combining different peptide identification methods within a parallel genetic algorithm. The distinctive feature of our approach, based on Pareto ranking, is that it can accommodate constraints and possibly conflicting scoring functions. We have also shown how population structure can significantly improve the wall clock time of a parallel peptide identification genetic algorithm while at the same time maintaining some exchange of information across local populations.

Improvements in performance modeling and identification of computational regimes within software libraries is a critical first step in developing software libraries that are truly agile with respect to the application as well as to the hardware. It is shown here that Pareto ranking, a concept from multi-objective optimization, can be an effective tool for mining large performance datasets. The approach is illustrated using software performance data gathered using both the public domain LAPACK library and an asynchronous communication library based on IBM LAPI active message library. Copyright © 2005 John Wiley & Sons, Ltd.

Restricted maximum likelihood (REML) estimation of variance–covariance matrices is an optimization problem that has both scientific and industrial applications. Parallel REML gradient algorithms are presented and compared for linear models whose covariance matrix is large, sparse and possibly unstructured. These algorithms are implemented using publicly available toolkits and demonstrate that REML estimates of large, sparse covariance matrices can be computed efficiently on multicomputers with hundreds of processors by using an effective mixture of data distributions together with a mixture of dense and sparse linear algebra kernels.

Gossiping is proposed as the preferred communication primitive for replicating pivot data in dense matrix factorization on message passing multicomputer. Performance gains are demonstrated on a hypercube for LU factorization algorithms based on gossiping as opposed to broadcasting. This finding has consequences for the design of numerical software libraries.

The cost of pivoting in LU factorization has become non-negligible on MIMD computers due to the acceleration of floating point arithmetic and comparatively slow communications. This study addresses the LU factorization of matrices stored by row. The importance of efficiently broadcasting pivot data is stressed. Multicasting is found preferable in this respect to broadcasting along minimum spanning trees. Threshold pivoting is shown to effectively reduce the number of messages while preserving a good load balance.