This is a multi-part message in MIME format. ------=_NextPart_000_0001_01BED6E2.B041AE40 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 8bit Hello, One more remark: instead of including MPI, we would better include some kind of numerical library with optimised implementations of some algorithms. These libraries can avoid or minimize among others cache trashing. They also give high performance implementations of BLAS level 1, 2 and 3. Most inexperienced programmers often reach only 1/3 of ½ of the maximum available computer power because of bad algorithm implementations. I think of libraries using MMX instead of the plain old x86 instructions. Using MMX (or its successor), WILL give you a big performance gain ! Multithreading will only give good performance in 10% of the cases. When applying a filter, communication overhead and synchronization will probably degrade any multithreading performance unless very experienced programmers wrote the filters AND if the images are large (> 10^6 pixels), large images also means large transports of data, so communication start-up time will be negligible. Bavo De Ridder bavodr@poboxes.com +32 (476) 24 58 36 ------=_NextPart_000_0001_01BED6E2.B041AE40 Content-Type: text/html; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable

Hello,

 

One more remark: instead of including MPI, we would better = include some kind of numerical library with optimised implementations of some = algorithms. These libraries can avoid or minimize among others cache trashing. They also = give high performance implementations of BLAS level 1, 2 and 3. Most = inexperienced programmers often reach only 1/3 of =BD of the maximum available = computer power because of bad algorithm implementations. I think of libraries using MMX instead of the plain old x86 instructions. Using MMX (or its successor), = WILL give you a big performance gain !

 

Multithreading will only give good performance in 10% of the = cases. When applying a filter, communication overhead and synchronization will = probably degrade any multithreading performance unless very experienced = programmers wrote the filters AND if the images are large (> 10^6 pixels), large = images also means large transports of data, so communication start-up time will = be negligible.

<= /p>

Bavo De Ridder<= /p>

bavodr@poboxes.com<= /p>

+32 (476) 24 58 = 36<= /p>

 <= /p>

------=_NextPart_000_0001_01BED6E2.B041AE40--