Student research opportunities
Parallel Linear Algebra Algorithms on the Intel Single-chip Cloud Computer
Project Code: CECS_654
This project is available at the following levels:
CS single semester, Honours, Summer Scholar, Masters
Keywords:
manycore computing, parallel computing, linear algebra, block-partitioned algorithms
Supervisor:
Dr Peter StrazdinsOutline:
Dense linear algebra computations are widely used in many parallel applications. The (High Performance) Linpack benchmark is of this nature; it is also (contentiously!) the most widely quoted supercomputer benchmark. Various parallelization strategies for distributed memory computers (i.e.. cluster computers) shave been developed for these algorithms.
The Intel Single-chip Cloud Computer (SCC) is a state-of-the-art experimental manycore chip multiprocessor. ON this chip, there are 48 cores. Unlike previous CMPs, these cores communicate using message passing rather than by shared memory. This is because the later requires cache coherency, which is not scalable with respect to energy or computation time with large number of cores. Thus, in terms of programming model and balance between compute and communicate speed, the SCC is very much like a cluster on a chip.
Goals of this project
This project will port parallel linear algebra libraries and algorithms to the SCC, and evaluate their relative performance. LU factorization will be of most interest, but Cholesky and QR factorization may also be considered. The strategies to be evaluated include storage blocking (HPL, DBLAS), algorithmic blocking (DBLAS), lookahead (HPL, DBLAS) and other load balancing techniques (Elemental, DBLAS).
Student Gain
Only a very few universities around the world have an SCC! This chip represents the future of manycore processing. The research is potentially highly publishable.
Background Literature
Start with the Optimal Load Balancing Techniques for Block-Cyclic Decompositions for Matrix Factorization and Lookahead and Algorithmic Blocking Techniques Compared for Parallel Matrix Factorization papers and their references from http://cs.anu.edu.au/~Peter.Strazdins/papers.
Links
Intel SCCHPL
Elemental
DBLAS
