Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.programming.threads > #2777 > unrolled thread

About my new algorithm

Back to article view | Back to comp.programming.threads

Contents

  About my new algorithm Ramine <ramine@1.1> - 2014-12-11 16:46 -0800

#2777 — About my new algorithm

Hello,

If you have noticed i have just come up with a new algorithm of
my Parallel conjugate gradient solver library that is NUMA aware,
this is really a big improvement over my previous algorithm,
my new algorithm contains two parst that are the most expensive,
and those teo parts are: a vector multiplication by a transpose of a 
matrix, and a vector multiplication by a matrix, but when i have 
parallelized my previous algorithm, i have parallelized just the memory 
data cache transfer from the L2 cache-line hit to the CPU that costs 
around 10 CPU cycles for every double type, and i have parallelized also 
the multiplication of two doubles and addition of two doubles,
but this was not enough, cause what we have to do also is parallelize
the memory data transfers from the memory to the L2 cache ,
and this is what we call a NUMA aware algorithm that really scale on
NUMA architecture, and this is what i have done in my new algorithm,
the memory data transfers from memory to the L2 cache was also
parallelized and this have made my new algorithm NUMA aware and really
scalable on NUMA architecture. But to become NUMA aware you need also 
to allocate memory of the arrays of your Matrix in different NUMA nodes, 
but that's easy to do.

My Parallel conjugate gradient solver library supports dense matrices
and it is a library that solves linear system of equations and also it 
solves large and very large dense linear system of equations.


You can download Parallel conjugate gradient solver library from:


https://sites.google.com/site/aminer68/parallel-implementation-of-conjugate-gradient-linear-system-solver


Thank you,
Amine Moulay Ramdane.

[toc] | [standalone]

Back to top | Article view | comp.programming.threads

csiph-web