Tunç Bahçecioğlu » coding

VS2010 good (sort of), vs2005 bad

tunc — Tue, 26 Oct 2010 13:43:06 +0000

I multiplied two 1000×1000 float matrices using the Eigen library on VS2005 and VS2010 with different optimizations, here are the results:

	VS2005 (s)	VS2010 ()
none	2.873	2.812
OpenMP	1.759	1.671
SSE	3.258	2.911
SSE2	5.057	0.224
OpenMP + SSE	1.807	1.598
OPENMP + SSE2	2.774	0.134

SO,

OpenMP support of Eigen works well.
Shouldn’t SSE work for floats ???
VS2005 is horrible with vectorization optimizations.
OPENMP + SSE2 + VS2010 with Eigen kicks ass!!!

MPI Debugging with Visual Studio

tunc — Mon, 17 Aug 2009 15:16:40 +0000

Quick article about debugging your parallel applications with Visual Studio

Open your project properties, and go to Configuration Properties->Debugging

All we have to do is to fill 3 textboxes:

MPIRun Command is the path + filename of your mpirun applcation(usually mpiexec.exe or mpirun.exe), In my case it is "c:\Program Files\MPICH2\bin\mpiexec.exe" (do not forget the “”).
MPIRun Arguments are the arguments we want to pass to the mPIRun command, at very least we have to tell mpi the number processes to spawn, so we use –n 2 for 2 processes (change 2 as you wish).
MPIShim Loacation is the path + filename of your mpishim.exe file. Mpishim.exe is diffrent for every release of visual studio and platform you have. For example its location for Visual Studio 2005 for 32bit platforms should be “C:\Program Files\Microsoft Visual Studio 8\Common7\IDE\Remote Debugger\x86”

delphi and mpi, closer look 2

tunc — Mon, 10 Aug 2009 11:11:42 +0000

It is pretty simple to install and use MPI with Delphi, fortunately for us programming the MPI is no bother too.

The second part is about calling MPI functions from Delphi.

Name Convention

No change here. Since we are using a direct translation of the original MPI header, all the original names are there.

Sending, Receiving Data

No surprise here too, here are some pointers:

Variables

Since MPI needs pointers to variables, use @ operator to get the pointer to your variables.

MPI_Comm_rank(MPI_COMM_WORLD,@myid);

For getting the size of variables use the sizeoff function.

Arrays

Again we need the starting address of the arrays.

Lets say we have an array named as “sizeArray”

For dynamic arrays use @(sizeArray[0])

For static arrays use @(sizeArray[1]) // assuming the array starts from 1

Strings

Since no one knows the memory map of a Delphi string you cannot use them with MPI. You have to go with short strings.

Debugging

Bad news here, Delphi has no support for parallel debugging. You have to go with displaying data to screen, or using a log file.

I was going to write a 3rd article about creating parallel applications with GUI, but since it is a general subject I decided to write the article separately, so this series and here.

DELPHI and MPI, closer look 1

tunc — Wed, 05 Aug 2009 13:39:45 +0000

Since my small article Delphi and MPI in 4 steps has its own fan base (almost everyday a coder lost in the world wide web visits this page), I decided to give the subject a closer look.

In this first part I will take a closer look at the installation process.

1. Install an MPI library

MPI itself is not a library but an interface. So you have to find an implementation suiting your needs. MPICH2 is a long time favorite of mine, as it is free, well developed and documented.

All you have to do download and install the library on every computer of your cluster.

2.Get a Binding

Unfortunately, MPICH2 does not include bindings for Delphi. Fortunately Lazarus people created a pascal file from the c include files that we can use.

You can find the slightly changed version of this file here Delphi mpi bindings (182). All you need to do is download the file and insert it into your project. Since the dll files from MPICH2 will be in a path where Windows can find no other action is necessary.

3.Code!

Add uses MPI; to your code, that’s it. Here is a small example:

program MPIExample;
uses  MPI;

var myid: integer;
begin
  MPI_Init(nil,nil);
  MPI_Comm_rank(MPI_COMM_WORLD,@myid);
  writeln('id=',myid);
  MPI_Finalize;

end.

pugg

tunc — Mon, 03 Aug 2009 00:15:30 +0000

a c++ framework for plug-in management

Features

Header only library, no installation
Auto loading of plug-ins from files
Version control for plug-ins
Object oriented design

Web Pages

http://pugg.sourceforge.net/
https://sourceforge.net/projects/pugg/

Installing boost for visual studio

tunc — Mon, 03 Aug 2009 00:06:01 +0000

Boost is a collection of free libraries for c++, kind of a Swiss knife for the c++ geek.
Most of boost is actually header only and don’t need any installation. To use the small dark part of boost
you need to install them, well built them, here is how:

Short, Softcore Method:

Go to BoostPro Computing and download a web installer. Follow the installer instructions.
It needs a free registration, and it does not contain some libraries like Boost.Python

Long, Hardcore Method:

This is the programmers way., we will built the libraries ourselves.

1.We start by downloading Boost Source code from Sourceforge

2.Extract the zip file to a nice folder, for me it is c:\Program Files\Boost\

3.Build BJam

Bjam is a tool that will help us built BOOST, but we first have to built it:
-Open visual studio Comman Prompt ,
start->All Programs->Microsoft Visual Studio 2005->visual studio tools->visual studio command prompt
cd C:\Program Files\boost\boost_1_39_0\tools\jam\src>
built.bat
This made the following C:\Program Files\boost\boost_1_39_0\tools\jam\src\bin.ntx86\bjam.exe
Now copy bjam.exe to the root boost directory : C:\Program Files\boost\boost_1_39_0\

4.Build Boost

from the visual studio command promt go to C:\Program Files\boost\boost_1_39_0\
For a complete set of libraries use the following:
bjam –toolset=msvc-8.0 –build-type=complete
If you want a sub set here are your options:
bjam toolset=msvc-8.0 variant=debug threading=multi link=shared
bjam toolset=msvc-8.0 variant=release threading=multi link=static
Now this will take some time…

5.Enjoy

Your lib directory is C:\Program Files\boost\boost_1_39_0\stage\lib
Yout include directory is C:\Program Files\boost\boost_1_39_0\include

For details of this process you can refer to Boost getting started web pages

A Comparative Study on Two Different Direct Parallel Solution Strategies for Large-Scale Problems

tunc — Sun, 02 Aug 2009 02:46:21 +0000

This is my latest publication

A Comparative Study on Two Different Direct Parallel Solution Strategies for Large-Scale Problems

T. Bahcecioglu, S. Ozmen and O. Kurc

This paper was presented in PARENG2009 Pecs, Hungary.
doi:10.4203/ccp.90.40
Web Page

Civil-Comp Proceedings

ISSN 1759-3433

CCP: 90

PROCEEDINGS OF THE FIRST INTERNATIONAL CONFERENCE ON PARALLEL, DISTRIBUTED AND GRID COMPUTING FOR ENGINEERING

Edited by: B.H.V. Topping and P. Iványi

Paper 40

A Comparative Study on Two Different Direct Parallel Solution Strategies for Large-Scale Problems

T. Bahcecioglu, S. Ozmen and O. Kurc

Department of Civil Engineering, Middle East Technical University, Ankara, Turkey

doi:10.4203/ccp.90.40

purchase the full-text of this paper

Full Bibliographic Reference for this paper

T. Bahcecioglu, S. Ozmen, O. Kurc, "A Comparative Study on Two Different Direct Parallel Solution Strategies for Large-Scale Problems", in B.H.V. Topping, P. Iványi, (Editors), "Proceedings of the First International Conference on Parallel, Distributed and Grid Computing for Engineering", Civil-Comp Press, Stirlingshire, UK, Paper 40, 2009. doi:10.4203/ccp.90.40

Keywords: parallel solution, multi-frontal, substructure, large-scale, PC cluster.

Summary

Parallel solution methods are being implemented in various commercial finite element analysis programs as a consequence of having large-scale structural models and increased affordability and availability of parallel computing environments. The current research on the parallel solution methods based on direct solution mainly focuses on two different kinds of solution strategies: global and domain-by-domain. Global solution strategies focus on the efficient solution of the system of linear equations. They require the entire stiffness matrix to be assembled and distributed to the computers. The domain-by-domain strategies, on the other hand, perform partitioning at the structural level and as a result they parallelize not only the equation solution but also the stiffness matrix generation, assembly, element force and stress computations.

This paper presents a comparative study on two different direct parallel solution strategies for the linear solution of large scale finite element models: global and domain-by-domain. The global solution strategy was examined by utilizing the parallel multi-frontal equation solver, MUMPS [1], together with a finite element program. In a similar manner a substructure based parallel solution framework [2] was utilized for investigating the domain-by-domain strategy.

Various large-scale structural models were solved with both global and domain-by-domain solution strategies in order to illustrate the efficiencies and weaknesses of each solution strategy. The test runs were performed using Windows XP running, a homogeneous PC cluster connected with an ordinary 1 GBit network switch.

The test runs performed for this study indicate that the global solution strategy was more efficient and faster than the domain-by-domain strategy for symmetric and uniform structural models but its efficiency was decreased as the symmetry and uniformity of the models diminished. The performance of the domain-by-domain strategy improved in non-uniform models and produced competitive solution times when compared to the times of the global solution strategy.

References

P.R. Amestoy, I.S. Duff, J.-Y. L’Excellent, "Multifrontal parallel distributed symmetric and unsymmetric Solvers", Comput. Methods Appl. Engrg., 184, 501-520, 2000. doi:10.1016/S0045-7825(99)00242-X
2: O. Kurc, "Parallel Computing in Structural Engineering", VDM Verlag, Germany, 2008.

send std::string with mpi

tunc — Sun, 02 Aug 2009 02:12:45 +0000

The idea is to send the size of the string first and then allocate the storage on the reciever. Here is an example with broadcast.

string s = "before";
int size = 0;
if (myRank == 0)
{
   s = "after";
   size = s.size();
}
BroadCast(&size,1);
if (myRank != 0) s.resize(size);
BroadCast(s.data(),size);

what i need from delphi 2010

tunc — Sun, 02 Aug 2009 01:55:51 +0000

64 bit compiler
OpenMP
Strict rules about data encapsulation
Getting rid of unit = namespace philosophy
Better compiler, fast is not always good enough!
Better editor, Visual Studio is faaar batter right now.

Delphi and mpi in 4 steps

tunc — Sun, 02 Aug 2009 01:31:28 +0000

Download and install MPICH2
Download mpi.pas from lazarus site and put it in your project. http://wiki.lazarus.freepascal.org/MPICH
Comment out any code that gives error in MPI.pas (Lazarus pascal is a little bit different than Delphi object pascal)
Write your code!

program HelloWorld1;

uses  MPI;

var myid: integer;
begin
MPI_Init(nil,nil);
MPI_Comm_rank(MPI_COMM_WORLD,@myid);
writeln('id=',myid);
MPI_Finalize;

end.