ITC Research Computing Support Using Matlab Effectively

1
ITC Research Computing SupportUsing Matlab
Effectively

• By Ed Hall
• Research Computing Support Center
• Phone [phone] F Fax [phone]
• E-Mail Res-Consult_at_Virginia.EDU
• http//www.itc.Virginia.edu/researchers

2
Matlab Overview
3
Matlab Overview
4
Matlab Overview
5
Matlab Overview
6
Matlab Overview

• UVA Matlab Resource webpage
• www.itc.virginia.edu/research/matlab/
• Matlab tutorial Getting Started
• www.mathworks.com/access/helpdesk/help/techdoc/l
  earn_matlab/learn_matlab.shtml

7
Matlab Programming
8
Matlab M-files
9
Matlab Function M-files
10
Matlab Function M-files
11
Matlab Function M-files
12
Maximizing MATLAB Performance

• Techniques for Improving Performance
• Performance Acceleration
• Sample Accelerated Program
• Measuring Performance
• Making Efficient Use of Memory

13
Techniques for Improving Performance

• Vectorizing loops

14
Vectorization (cont.)

• Logical indexing of arrays

15
Vectorization (cont.)
16
Preallocating Arrays

• Avoids overhead of dynamic resizing
• Reduces memory fragmentation
• Use array appropriate preallocation

17
Preallocation (cont.)

• Preallocating a Nondouble Matrix
• Use repmat to enlarge arrays

18
Other Ways to Speed Performance

• Coding loops in a MEX-file for speed using Matlab
  compiler
• Functions M-files are faster than scripts
• Load and Save are faster than File I/O Functions
• Avoid large background processes

19
Performance Acceleration

• JIT-Accelerator, new in Matlab 6.5
• Make MATLAB as fast as C and Fortran
• Most improvement in self-contained loops
• As fast as vectorized code
• Improvement may vary depending on code

20
What MATLAB Accelerates

• Performance enhancement supported on subset of
  Matlab

21
What Matlab Accelerates (cont.)

• Arrays shapes of lt 3 dimensions
• for loops faster as long as
• Conditional statements (if, elseif, while,
  switch) that evaluate to a scalar
• Overhead of manipulating small arrays

22
What MATLAB Doesnt Accelerate

• Some data types and array shapes
• Calls to other functions (M-File of MEX)
• More than one operation on a line
• First instruction disqualifies entire line

23
What MATLAB Doesnt Accelerate

• Changing data types or variable shapes
• Using complex constants in scripts

24
Perfomance Acceleration (cont.)

• How Vectorization and Preallocation fit in
• Acceleration functionally equivalent to
  vectorization
• Choose method most understandable
• Preallocation still important
• Avoid running in debug mode
• Aviod command echoing

25
Measuring Performance

• Profiling measure where program spends the most
  time
• Create simple implementation dealing with
  obvious performance issues
• Use profiler to identify performance bottlenecks
  (80/20 rule)
• Identify time-costly functions and where they are
  called

26
Measuring Performance (cont.)

• Problems identified by profiler solved by
• Avoiding unnecessary computation
• Changing algorithm to avoid time-costly
  functions
• Avoiding recomputation by storing results for
  future use
• Want most time spent on calls to small number of
  builtin functions

27
Matlab Profiler

• GUI interface to profile function
• See Matlab News Notes example
• www.mathworks.com/company/newsletter/may03/pr
  ofiler.shtml
• Process for Improving Performance

28
Matlab Profiler (cont.)
29
Matlab Profile Function

• Command line profiler
• gtgtprofile on detail builtin history
• gtgt(Execute your M-file)
• gtgtprofile report
• gtgtprofile off
• Some features not in GUI Profiler

30
Making Efficient Use of Memory

• Memory management functions

31

• Memory Management Functions (cont.)

32
Ways to Conserve Memory
33
Ways to Conserve Memory

• Compressing data in memory
• Memory fragmentaion prevents contiguous memory
  block to store new large variable
• Out of Memory message
• Pack function compresses data in memory,
  freeing large contiguous blocks

34
Ways to Conserve Memory
35
Unix-Specific Memory Topics

• Freeing clear memory on Unix
• Matlab does not free up cleared memory in Unix
  until it Matlab exits
• Save workspace variables
• Exit Matlab
• Startup Matlab
• Load variables back in new workspace

36
Unix-Specific Memory Topics

• Additional Memory used to execute external
  commands
• Out of Memory error when executing Unix
  command from within Matlab (using shell escape
  (!) operator)
• Unix allocates as much memory to subprocess as
  parent process

37
Unix-Specific Memory Topics

• Reusing Heap Memory
• Matlab request memory from Unix when not
  enough available in Matlab heap

38
Matlab Compiler

• Converts function M-files to C or C source code
• Uses native compiler to create executables
• Creates MEX-file from function M-file, e.g.
  function.m
• gtgtmcc x function
• Creates stand-alone executable from function
  M-file
• gtgtmcc m function

39
Matlab Compiler
40
Matlab Compiler
41
Matlab Compiler
42
Unix-Specific Compiler Topics

• Setting LD_LIBRARY_PATH (e.g. in .kshrc) to
  locate shared libraries for stand-alone
  executables

43
Running Matlab in Batch Mode

• On a Unix platform

44
Further Resources

• Using Matlab Programming and Data Types
• www.mathworks.com/access/helpdesk/help/techdoc/mat
  lab_prog/matlab_prog.shtml
• Code Vectorization Guide
• www.mathworks.com/support/tech-notes/v5/1100/11
  09.shtml
• Matlab Compiler Users Guide
• www.mathworks.com/access/helpdesk/help/pdf_doc/co
  mpiler/compiler3.pdf