Jesmin Jahan PhD Defense

Event Type: 
Faculty Candidate
Dates: 
Wednesday, December 2, 2015 - 11:30 to 13:50
Location: 
120 Conference Room
Title: Engineering High-performance Parallel Algorithms with Applications to Bioinformatics
 
Abstract:
 
Since the beginning of the last decade, plateauing of the clock speed of computer processors has forced us to invest more in parallelism — for both hardware and software. This resulted in improvements in computing technologies that have favored parallelism over increased clock speed. Taking advantage of these improvements requires designing algorithms that can leverage parallelism well. In this dissertation, we show how to take advantage of several algorithm design techniques to harness modern heterogeneous parallel architectures for solving problems in bioinformatics efficiently. Our main goal while designing algorithms is to achieve high-performance in terms of running time and scalability. Other desirable goals include energy efficiency, portability and adaptivity.
 
We solve bioinformatics problems on grids (dynamic programming problems), on graphs (breadth-first search), and problems that can be solved using spatial trees (Molecular Dynamics using octrees). We present many novel algorithms, algorithm engineering techniques, theoretical analyses and performance evaluations on a range of state-of-the-art parallel architectures including multicores, manycores, and special purpose accelerators. Although we mainly target problems in bioinformatics, the algorithmic techniques we use to solve those problems have general applicability.
 
For many dynamic programming problems, we show that if we use a cache-oblivious recursive divide-and-conquer technique to solve them, the resulting algorithms become highly optimizable, cache-optimal and often have asymptotically better parallelism than their standard iterative counterparts. These algorithms not only have good theoretical bounds, but also perform better than standard iterative and tiled-loop algorithms in terms of running time, scalability, energy-efficiency, cache-adaptivity and portability. Furthermore, it is often possible to improve parallelism of these recursive algorithms without sacrificing cache-optimality by removing artificial dependencies among the tasks introduced by the recursive structure of the algorithm.
 
Breadth first search is a popular graph traversal algorithm that has many applications in bioinformatics. We show how to use lock- and atomic-instruction-free optimistic parallelization to improve parallelism and load balancing in a shared-memory parallel breadth-first search (BFS) algorithm. We present several work-efficient parallel BFS algorithms (including one that uses recursive divide-and-conquer) along with their theoretical and empirical performance analyses on state-of-the art multicore and manycore architectures.
 
Spatial trees (e.g., quad tree, octree, k-D tree) are recursive space partitioning data structures that are often used to store biological molecules efficiently. We present octree-based distributed and distributed-shared-memory hybrid near-far approximation algorithms to compute molecular polarization energy. These algorithms outperform all other state-of-the-art Molecular Dynamics packages by orders of magnitude on multicores and clusters of multicores.
 
We conclude by discussing implications of our work for future parallel algorithm design, and ways to extend our research to other domains.
Hosted By: 
Rezaul Chowdhury
Computed Event Type: 
Mis
Event Title: 
Jesmin Jahan PhD Defense