BMI 6420 Lecture

1
BMI 6420Lecture 1

• Julio C. Facelli
• Department of Biomedical Informatics
• and
• Center for High Performance Computing
• University of Utah
• Julio.Facelli_at_utah.edu

2
Discussion of syllabus
3
HPC Taxonomy

• Vector Supercomputers
• RISC
• Commodity
• New architectures (FPGA, Cell, GPUs)

4
John D. McCalpin, http//www.cs.virginia.edu/strea
m/
5
(No Transcript)
6
(No Transcript)
7
Parallel Architectures

• SMP
• cc-NUMA
• Clusters

8
SMP
9
cc-NUMA
10
CLUSTER
11
Comparison of Architectures
12
Parallel Computing

• Given one fast processor, N processors are
  faster.
• N processors of speed1 are cheaper than One
  processor of speed N.
• Efficiency is inversely proportional to number of
  processors.

13
Conclusions I

• Parallel (scientific) computing is unavoidable
• Memory locality is important
• Performance highly dependent of matching between
  algorithm and architecture

14
Top 500

• http//www.top500.org
• (List Charts)
• http//www.netlib.org/utk/papers/advanced-computer
  s/

15
FPGAs

• http//www.xilinx.com/
• http//www.ncsa.uiuc.edu/pointer/whitepapers.html
  

16
Cell Processors

• http//www-304.ibm.com/jct03004c/businesscenter/ve
  nturedevelopment/us/en/featurearticle/gcl_xmlid/86
  49/nav_id/emerging

17
GPUs

• http//www.nvidia.com/page/technology_vista_home.h
  tml

18
Parallel Computing Programming

• Data distributed model
• C, C, FORTRAN
• plus
• MPI (http//www-unix.mcs.anl.gov/mpi/)

19
Parallel Computing Tutorial

• http//www.llnl.gov/computing/tutorials/parallel_c
  omp/
• http//www.cs.rit.edu/ncs/parallel.htmlcourses

20
Granularity

• Independent execution blocks
• Low
• High