[Insight-users] Guidelines for speed and memory footprint optimization

[Matt McCormick]

Sergio Vera <sergio.vera at ...> writes:

> 
> 
> Also I may ask if you use streaming when there is no writer at the end? and if 
yes, how you enable streaming in this case?
> Thanks

You can use the following filter
http://www.itk.org/Doxygen320/html/classitk_1_1StreamingImageFilter.html#_detail
s


> 
> I have identified several ways for optimizing my code, but I was wondering if 
there are other things I could do. Here is what I have found until now:
> 
> 1. For multi-channel images, the use of scalar images with FixedArray pixels 
gives a 1.5 speedup factor with respect to the use of VectorImage with scalar 
pixel types (I know the number of bands at compilation time).
> 

Interesting.

> 
> 2. Even though I need to work with real types for the computations (physical 
magnitudes between 0 and 1), I found that writing intermediate results to files 
using real types is much slower than multiplying the values by a large constant 
and writing the result to disk using integer types.
> 

Reading/writing to disk can be a bottleneck.  Depending on the situation, 
writing to a tmpfs can be helpful.
http://en.wikipedia.org/wiki/Tmpfs

> 
> 
> I was also wondering if there is a smart way to choose the number of streaming 
divisions for the writer in order to ensure that the memory footprint of my 
program does not goes beyond a given size (other than try and error).
> 

Profiling software not specific to ITK can be helpful for both decreasing run 
times and memory footprints.  My preference is valgrind.
http://valgrind.org/

For determining bootlenecks, try the valgrind callgrind tool + kcachegrind.
http://kcachegrind.sourceforge.net/

For memory profiling, try the valgrind massif tool + ms_print that comes with 
valgrind.

> 
> PS: If there are general guidelines for this kind of optimization, maybe there 
should be a chapter in the software guide or a wiki page for this topic?
> 

A wiki article would be cool.

Regards,
Matt