NAIS Postgraduate Researchers

Bubacarr Bah School of Mathematics, University of Edinburgh JCMB 5620, King's Buildings Edinburgh, EH9 3JZ +44 (0) 131 650 5083 b.bah@sms.ed.ac.uk

I am interested in random matrix questions arising in signal processing particularly in compressed sensing and related areas like sparse approximation and matrix completion. Most of the analysis in these areas make use a random matrix quantity, restricted isometry constant (RIC) of the measurement matrix. Due to the intractability of RICs or RIC bounds for deterministic matrices but the possibility of probabilistic RIC bounds, I work on deriving and numerically computing sharp bounds for both dense, [1], and sparse random matrix ensembles. [1]B. Bah and J. Tanner, Improved Bounds on Restricted Isometry Constants for Gaussian Matrices, SIAM J. on Matrix Analysis, vol. 31, no. 5, pp. 2882-2898 (2010)

Alexander Collins 10 Crichton Street Edinburgh EH8 9AB A.J.Collins-2@sms.ed.ac.uk

I am interested in auto-tuning parallel programs, through the use of parallel skeletons and machine learning, in order to maximise program performance. This is applied to a variety of architectures, including shared-memory multi-core systems and GPGPU.

Qi Huangfu Room 5406 JCMB The King's Buildings, Mayfield Road Edinburgh s0789844@sms.ed.ac.uk

My major research area is simplex technique and its implementation. Simplex method (since 1940s) is a popular algorithm for linear programming (LP). Our research target is a parallel implementation of the simplex method that gives good speed-up on general, large sparse LP problems. We have got some interesting initial results. [1]J.A.J. Hall and Q. Huangfu, A high performance dual revised simplex solver, Technical Report ERGO 11-007, School of Mathematics

Issa Karambal Department of Mathematics Heriot-Watt University ik68@hw.ac.uk

My research interest lies in the numerical computation of the spectrum of a non-selfadjoint operator arising from the travelling wave problem. The numerical and theoritical approach used to investigate the problem is the Evans function (the shooting method ) and the Fredholm determinant. Also we investigate the connection between these two functions.

Charles Matthews Room 5406, JCMB King's Buildings, Edinburgh EH9 3JZ charles.matthews@ed.ac.uk

I am a PhD student at Edinburgh University with a broad interest in applied mathematics but particularly in efficient methods for the integration of (stochastic) differential equations related to Molecular Dynamics. My focus is in practical results that provide useful information on computation or implementation of Molecular Dynamics problems. My current work is looking at numerical methods for integrating stochastic perturbations to Newton's equations in order to sample the canonical (Gibbs) distribution, which corresponds to "constant-temperature Molecular Dynamics". This can be done in multiple ways, but commonly involves splitting up the vector field and integrating each piece separately. Generally a symmetric approach is used, but in terms of the order of terms integrated, and how the vector field is split, there is little "good" advice in the literature. Currently I am looking at how sampling is helped or hindered by shuffling these terms. Additionally I have worked with Accelrys on their Material Studio molecular dynamics package, implementing some new algorithms for numerical integration. Current work involving the "correct" parametrisation of this method and a study of using it to thermostat "real" systems with a very large number of degrees of freedom is underway.

Andrew McPherson 1.05 Informatics Forum 10 Crichton Street EH8 9AB + 044 131 650 5146 ajmcpherson@ed.ac.uk

Currently investigating the compile time analysis and transformation of MPI programs, to optimise their communication on multicore clusters. The aim being to build a static understanding of the communication characteristics of an MPI program. This is then used to optimise scheduling and enable the use of shared memory communication where possible.

Emani Murali Room 1.05, School of Informatics 10 Crichton Street Edinburgh EH89AB (UK) m.k.emani@sms.ed.ac.uk

My research is on developing a "Predictive Modelling based approach to Runtime adaptation of parallel programs". We try to build a model where the mapping of parallel programs is predicted at runtime where there are external varying workloads in the environment.

Mark Payne Department of Mathematics Heriot-Watt University Edinburgh mjp8@hw.ac.uk

Research in numerical analysis, in particular developing and adapting finite element methods for time dependent wave propogation problems. This includes comparing previously developed techniques, producing methods of mass matrix diagonalisation and methods with high order accuracy. Work will lead to implementation of techniques on high performance parallel computers.

Martin Takac The University of Edinburgh James Clerk Maxwell Building Mayfield Road Edinburgh Scotland EH9 3JZ takac.mt@gmail.com

Parallel coordinate descent method for huge-scale problems, Iteration complexity of algorithms for convex optimization [1]P. Richt\'{a}rik and M. Tak\'{a}\v{c}: Efficient serial and parallel coordinate descent methods for huge-scale truss topology design, ERGO Technical Report 11-012 (2011) [2] P. Richt\'{a}rik and M. Tak\'{a}\v{c}: Efficiency of randomized coordinate descent methods on minimization problems with a composite objective function, SPARS 11 Proceedings (2011) [3]P. Richt\'{a}rik and M. Tak\'{a}\v{c}: Iteration complexity of randomized block-coordinate descent methods for minimizing a composite function, ERGO Technical Report 11-011 (2011)

Luis Fabricio Wanderley Goes University of Edinburgh Informatics Forum 10 Crichton Street Edinburgh EH8 9AB 131 651 5661 lfwgoes.inf@ed.ac.uk

The recent shift toward multi-core chips has pushed the burden of extracting performance to the programmer. In fact, programmers now have to be able to uncover more coarse-grain parallelism with every new generation of processors, or the performance of their applications will remain roughly the same or even degrade. Unfortunately, parallel programming is still hard and error prone. This has driven the emergence of many new programming models that aim to make this process efficient. Transactional Memory is an attracting alternative parallel programming model. From a different perspective, it simplifies parallel programming by removing the burden of correctly synchronizing threads on data races. This model allows programmers to write parallel code as transactions, which are then guaranteed by the runtime system to execute atomically and in isolation regardless of eventual data races. Although removing the burden of correctly synchronizing parallel applications is an important simplification, the programmer is still left with the tasks of thread scheduling and orchestration. These tasks can be naturally handled by skeleton or pattern-based programming. It allows parallel programs to be expressed as specialized instances of generic communication and computation patterns. This leaves the programmer with only the implementation of the particular operations required to solve the problem at hand. Thus, this programming approach eliminates some of the major challenges of parallel programming, namely thread communication and orchestration. In addition to simplifying the programming task, skeletons are also amenable to performance optimizations. I am currently working on a new skeleton framework that selects and applies performance optimizations in transactional worklist applications. It uses a novel hierarchical autotuning mechanism that dynamically selects the most suitable set of optimizations for each application and adjusts them accordingly. Additionally, I am also investigating the performance impact of existing system-level optimizations when applied to transactional worklist applications.

Cheherazada Gonzalez Dept. Of Mathematics and Statistics, University of Strathclyde 26, Richmond Street Glasgow G1 1XH

Heather Yorston Dept. Of Mathematics and Statistics, University of Strathclyde 26, Richmond Street Glasgow G1 1XH