CS 267: Applications of Parallel Computers Final Project Suggestions

1
CS 267 Applications of Parallel
ComputersFinal Project Suggestions

• James Demmel
• www.cs.berkeley.edu/demmel/cs267_Spr06

2
Outline

• Kinds of projects
• Evaluating and improving the performance of a
  parallel application
• Application could be full scientific
  application, or important kernel
• Parallelizing a sequential application
• other kinds of performance improvements possible
  too, eg memory hierarchy tuning
• Devise a new parallel algorithm for some problem
• Porting parallel application or systems software
  to new architecture
• Example of previous projects (all on-line)
• Upcoming guest lecturers
• See their previous lectures, or contact them, for
  project ideas
• Suggested projects

3
CS267 Class Projects from 2004

• BLAST Implementation on BEE2  Chen Chang
• PFLAMELET An Unsteady Flamelet Solver for
  Parallel Computers  Fabrizio Bisetti
• Parallel Pattern Matcher  Frank Gennari, Shariq
  Rizvi, and Guille Díez-Cañas
• Parallel Simulation in Metropolis  Guang Yang
• A Survey of Performance Optimizations for
  Titanium Immersed Boundary Simulation  Hormozd
  Gahvari, Omair Kamil, Benjamin Lee, Meling Ngo,
  and Armando Solar
• Parallelization of oopd1  Jeff Hammel
• Optimization and Evaluation of a Titanium
  Adaptive Mesh Refinement Code  Amir Kamil, Ben
  Schwarz, and Jimmy Su

4
CS267 Class Projects from 2004 (cont)

• Communication Savings With Ghost Cell Expansion
  For Domain Decompositions Of Finite Difference
  Grids  C. Zambrana Rojas and Mark Hoemmen
• Parallelization of Phylogenetic Tree
  Construction  Michael Tung
• UPC Implementation of the Sparse Triangular Solve
  and NAS FT  Christian Bell and Rajesh Nishtala
• Widescale Load Balanced Shared Memory Model for
  Parallel Computing  Sonesh Surana, Yatish Patel,
  and Dan Adkins

5
Planned Guest Lecturers

• Katherine Yelick (UPC, heart modeling)
• David Anderson (volunteer computing)
• Kimmen Sjolander (phylogenetic analysis of
  proteins SATCHMO Bonnie Kirkpatrick)
• Julian Borrill, (astrophysical data analysis)
• Wes Bethel, (graphics and data visualization)
• Phil Colella, (adaptive mesh refinement)
• David Skinner, (tools for scaling up
  applications)
• Xiaoye Li, (sparse linear algebra)
• Osni Marques and Tony Drummond, (ACTS Toolkit)
• Andrew Canning (computational neuroscience)
• Michael Wehner (climate modeling)

6
Suggested projects (1)

• Weekly research group meetings on these and
  related topics (see J. Demmel and K. Yelick)
• Contribute to upcoming ScaLAPACK release (JD)
• Proposal, talk at www.cs.berkeley.edu/demmel
  ask me for latest
• Performance evaluation of existing parallel
  algorithms
• Ex New eigensolvers based on successive band
  reduction
• Improved implementations of existing parallel
  algorithms
• Ex Use UPC to overlap communication, computation
• Many serial algorithms to be parallelized
• See following slides

7
Missing Drivers in Sca/LAPACK
LAPACK ScaLAPACK
Linear Equations LU Cholesky LDLT xGESV xPOSV xSYSV PxGESV PxPOSV missing
Least Squares (LS) QR QRpivot SVD/QR SVD/DC SVD/MRRR QR iterative refine. xGELS xGELSY xGELSS xGELSD missing missing PxGELS missing missing missing (intent?) missing missing
Generalized LS LS equality constr. Generalized LM Above Iterative ref. xGGLSE xGGGLM missing missing missing missing
8
More missing drivers
LAPACK ScaLAPACK
Symmetric EVD QR / BisectionInvit DC MRRR xSYEV / X xSYEVD xSYEVR PxSYEV / X PxSYEVD missing
Nonsymmetric EVD Schur form Vectors too xGEES / X xGEEV /X missing driver missing driver
SVD QR DC MRRR Jacobi xGESVD xGESDD missing missing PxGESVD missing (intent?) missing Missing
Generalized Symmetric EVD QR / BisectionInvit DC MRRR xSYGV / X xSYGVD missing PxSYGV / X missing (intent?) missing
Generalized Nonsymmetric EVD Schur form Vectors too xGGES / X xGGEV / X missing missing
Generalized SVD Kogbetliantz MRRR xGGSVD missing missing (intent) missing
9
Suggested projects (2)

• Contribute to sparse linear algebra (JD KY)
• Performance tuning to minimize latency and
  bandwidth costs, both to memory and between
  processors (sparse gt few flops per memory
  reference or word communicated)
• Typical methods (eg CG conjugate gradient) do
  some number of dot projects, saxpys for each
  SpMV, so communication cost is O( iterations)
• Our goal Make latency cost O(1)!
• Requires reorganizing algorithms drastically,
  including replacing SpMV by new kernel Ax, A2x,
  A3x, , Akx, which can be done with O(1)
  messages
• Projects
• Study scalability bottlenecks of current CG on
  real, large matrices
• Optimize Ax, A2x, A3x, , Akx on sequential
  machines
• Optimize Ax, A2x, A3x, , Akx on parallel
  machines

10
Suggested projects (3)

• Evaluate new languages on applications (KY)
• UPC or Titanium
• UPC for asynchrony, overlapping communication
  computation
• ScaLAPACK in UPC
• Use UPC-based 3D FFT in your application
• Optimize existing 1D FFT in UPC, to use 3D
  techniques
• Porting, Evaluating parallel systems software
  (KY)
• Port UPC to RAMP
• Port GASNET to Blue Gene, evaluate performance