Saliya's Blogs

Get Process Affinity

#define _GNU_SOURCE
#include 
#include 
#include 
#include 
#include 

int main(int argc, char* argvi[])
{
    pid_t pid = getpid();
    cpu_set_t my_set;
    int ret;

    CPU_ZERO(&my_set);
    ret = sched_getaffinity(0, sizeof(my_set), &my_set);
    char str[80];
    strcpy(str," ");
    int count = 0;
    int j;
    for (j = 0; j < CPU_SETSIZE; ++j)
    {
        if (CPU_ISSET(j, &my_set))
        {
            ++count;
            char cpunum[3];
            sprintf(cpunum, "%d ", j);
            strcat(str, cpunum);
        }
    }
    printf("pid %d affinity has %d CPUs ... %s\n", pid, count, str);
    return 0;
}

taskset -c 1,2 ./a.out

Set Process Affinity

#define _GNU_SOURCE
#include 
#include 
#include 
#include 
#include 

int main(int argc, char* argvi[])
{
    pid_t pid = getpid();
    cpu_set_t my_set;
    int ret;

    CPU_ZERO(&my_set);
    CPU_SET(1, &my_set);
    CPU_SET(2, &my_set);

    ret = sched_setaffinity(0, sizeof(my_set), &my_set);
    printf("pid %d \n", pid);

    // A busy loop to keep the program from terminating while
    // you use taskset to check if affinity is set as you wanted
    long x = 0;
    long i;
    while (i < LONG_MAX)
    {
        x += i;
        ++i;
    }
    printf("%ld\n",x);
    return 0;
}

taskset -pc pid

If you encounter an error similar to the following, which complains that GLIBCXX_3.4.9 could not be found, while running an application with Apache Spark you can avoid this by switching Spark's compression method from snappy to something such aslzf.

...
Caused by: java.lang.UnsatisfiedLinkError: .../snappy-1.0.5.3-1e2f59f6-8ea3-4c03-87fe-dcf4fa75ba6c-libsnappyjava.so: /usr/lib64/libstdc++.so.6: version `GLIBCXX_3.4.9' not found (required by.../snappy-1.0.5.3-1e2f59f6-8ea3-4c03-87fe-dcf4fa75ba6c-libsnappyjava.so)

There are a few ways how one can pass configuration options to Spark. The naive way seems to be through command line as,

--conf "spark.io.compression.codec=lzf"

On a side note, you can find what GLIBC versions are available by running strings /usr/lib/libstdc++.so.6 | grep GLIBC

References

The ability to refer (pass) methods in Java 8 is a convenient feature, however, as a programmer you might face the situtaion where some code that seemingly follow the correct syntax to refer a method that throws an exception gives a compilation error of anUnhandled Exception, which doesn't go away by wrapping the call in a try/catch or adding a throws clause to the method signature. See the following code,

import java.util.function.Function;

public class PassingMethodsThatThrowExceptions {

    public static int addOne(String value) throws NotANumberException{
        int v = 0;
        try{
            v = Integer.parseInt(value);
        } catch (NumberFormatException e){
            throw new NotANumberException();
        }
        return v+1;
    }

    public static void increment(Function<String,Integer> incrementer, String value){
        System.out.println(incrementer.apply(value));
    }

    public static void main(String[] args) {
        increment(PassingMethodsThatThrowExceptions::addOne, "10");
    }
}

This is a simple code, which has

Note. addOne function is declared to throw possible exception of type NotANumberException (the type of exception is NOT important here).

This code will result in following compilation error,

    Error: java: incompatible thrown types exceptions.NotANumberException in method reference

If you use an IDE such as IntelliJIDEA it'll show Unhandled Exception: NotANumberException for the increment method call in mainand adding try/catch will not work.

What's going wrong here? It's actually a mistake on your end.

The increment function expects a function that takes a String and returns an int, but you forgot to mention that this method may also throw an exception of type NotANumberException.

The solution is to correct the type of incrementer parameter in increment function.

Note. you'll need to write a new functional interface because you can't add throws NotANumberException to thejava.util.function.Function interface that's used to define the type of incrementer parameter here.

Here's the working solution in full.

public class PassingMethodsThatThrowExceptions {
    public interface IncrementerSignature{
        public int apply(String value) throws NotANumberException;
    }

    public static int addOne(String value) throws NotANumberException{
        int v = 0;
        try{
            v = Integer.parseInt(value);
        } catch (NumberFormatException e){
            throw new NotANumberException();
        }
        return v+1;
    }

    public static void increment(IncrementerSignature incrementer, String value) throws NotANumberException {
        System.out.println(incrementer.apply(value));
    }

    public static void main(String[] args) {
        try {
            increment(PassingMethodsThatThrowExceptions::addOne, "10");
        } catch (NotANumberException e) {
            e.printStackTrace();
        }
    }
}

Also, note this is NOT something to do with referring methods or Java 8 in general. You may face a similar situation even in a case where you implement a method of an interface and in the implementation you add the throws SomeException to the signature. Here's a stackoverflow post you'd like to see on this.

Hope this helps!

tl;dr - Use Dillinger and paste the formatted content directly to blogger

Recently, I tried many techinques, which will allow me to write blogs in markdown. The available choice in broad categories are,

First is the obvious choice if you need total control over your blog, but I didn't want to get into too much trouble just to blog because it involes hosting the generated static html pages on your own - not to mention the trouble of enabling comments. I like the second solution from and went the distance to even move my blog to svbtle. It's pretty simple and straightforward, but after doing a post or two I realized the lack of comments is a showstopper. I agree it's good for posts intended for "read only" use, but usually it's not the case for me.

This is when I started investigating on the third option and thought StackEdit to be a nice solution as it'll allow posting to blogger directly. However, it doesn't support syntax highlighting for code blocks - bummer!

Then came the "aha!" moment. I've been using Dillinger to edit markdown regularly as it's very simple and gives you instant formatted output. I thought why not just copy the formatted content and paste it in the blog post - duh. No surprises - it worked like a charm. Dillinger beatifully formats everything including syntax highligting for code/scripts. Also, it allows you to link with either Dropbox or Github where I use Github.

All in all, I found Dillinger to be the easiest solution and if you like to see a formatted post see my first post with it.

My 10 month old son is pioneering his crawling skills and has just begun to cruise. It's been hard to keep him out of the shoe rack with these mobile skills, so I decided to make this little fence.

Download Sketchup file

Download PDF file

Here's a video of the sliding lock mechanism I made.

• Mono 3.4.0
• MPI.NET source code revision 338.

    svn co https://svn.osl.iu.edu/svn/mpi_net/trunk -r 338 mpi.net
  

  • Also, download the Unsafe.pl.patch, which was originally available here
• OpenMPI 1.4.3. Note this is a retired version of OpenMPI and we are using it only because that's the best that I could get MPI.NET to compile against. If in future MPI.NET team provides support for a newer version of OpenMPI, you may be able to use it as well.
• Automake 1.9. Newer versions may work, but I encountered some errors in the past, which made me stick with version 1.9.

1. I suggest installing everything to a user directory, which will avoid you requiring super user privileges. Let's create a directory called build_mono inside home directory.

    mkdir ~/build_mono
   

   The following lines added to your ~/.bashrc will help you follow the rest of the document.

    BUILD_MONO=~/build_mono
    PATH=$BUILD_MONO/bin:$PATH
    LD_LIBRARY_PATH=$BUILD_MONO/lib
    ac_cv_path_ILASM=$BUILD_MONO/bin/ilasm
   
    export BUILD_MONO PATH LD_LIBRARY_PATH ac_cv_path_ILASM
   

   Once these lines are added do,

    source ~/.bashrc
   

2. Build automake by first going to the directory that containst automake-1.9.tar.gz and doing,

    tar -xzf automake-1.9.tar.gz
    cd automake-1.9
    ./configure --prefix=$BUILD_MONO
    make
    make install
   

   You can verify the installation by typing which automake, which should point to automake inside $BUILD_MONO/bin
3. Build OpenMPI. Again, change directory to where you downloaded openmpi-1.4.3.tar.gz and do,

    tar -xzf openmpi-1.4.3.tar.gz
    cd openmpi-1.4.3
    ./configure --prefix=$BUILD_MONO
    make
    make install
   

   Optionally if Infiniband is available you can point to the verbs.h (usually this is in /usr/include/infiniband/) by specifying the folder /usr in the above configure command as,

    ./configure --prefix=$BUILD_MONO --with-openib=/usr
   

   If building OpenMPI is successfull, you'll see the following output for mpirun --version command,

    mpirun (Open MPI) 1.4.3
   
    Report bugs to http://www.open-mpi.org/community/help/
   

   Also, to make sure the Infiniband module is built correctly (if specified) you can do,

    ompi_info|grep openib
   

   which, should output the following.

    MCA btl: openib (MCA v2.0, API v2.0, Component v1.4.3)
   

4. Build Mono. Go to directory containing mono-3.4.0.tar.bz2 and do,

    tar -xjf mono-3.4.0.tar.bz2
    cd mono-3.4.0
   

   Mono 3.4.0 release is missing a file, which you'll need to add by pasting the following content to a file called./mcs/tools/xbuild/targets/Microsoft.Portable.Common.targets

    <Project xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
        <Import Project="..\Microsoft.Portable.Core.props" />
        <Import Project="..\Microsoft.Portable.Core.targets" />
    </Project>
   

   You can continue with the build by following,

    ./configure --prefix=$BUILD_MONO
    make
    make install
   

   There are several configuration parameters that you can play with and I suggest going through them either in README.md or in./configure --help. One parameter, in particular, that I'd like to test with is --with-tls=pthread
5. Build MPI.NET. If you were wonder why we had that ac_cv_path_ILASM variable in ~/.bashrc then this is where it'll be used. MPI.NET by default tries to find the Intermediate Language Assembler (ILASM) at /usr/bin/ilasm2, which for 1. does not exist because we built Mono into $BUILD_MONO and not /usr 2. does not exist because newer versions of Mono calls this ilasm notilasm2. Therefore, after digging through the configure file I found that we can specify the path to the ILASM by exporting the above environment variable.
   Alright, back to building MPI.NET. First copy the downloaded Unsafe.pl.patch to the subversion checkout of MPI.NET. Then change directory there and do,

    patch MPI/Unsafe.pl < Unsafe.pl.patch
   

   This will say some hunks failed to apply, but that should be fine. It only means that those are already fixed in the checkout. Once patching is completed continue with the following.

    ./autogen.sh
    ./configure --prefix=$BUILD_MONO
    make
    make install
   

   At this point you should be able to find MPI.dll and MPI.dll.config inside MPI directory, which you can use to bind against your C# MPI application.

• Here's a sample MPI program written in C# using MPI.NET.

    using System;
    using MPI;
    namespace MPINETinMono
    {
        class Program
        {
            static void Main(string[] args)
            {
                using (new MPI.Environment(ref args))
                {
                    Console.Write("Rank {0} of {1} running on {2}\n",
                                    Communicator.world.Rank,
                                    Communicator.world.Size,
                                    MPI.Environment.ProcessorName);
                }
            }
        }
    }
  

• There are two ways that you can compile this program.
  1. Use Visual Studio referring to MPI.dll built on Windows
  2. Use mcs from Linux referring to MPI.dll built on Linux

     mcs Program.cs -reference:$MPI.NET_DIR/tools/mpi_net/MPI/MPI.dll
     

     where $MPI.NET_DIR refers to the subversion checkout directory of MPI.NET
     Either way you should be able to get Program.exe in the end.
• Once you have the executable you can use mono with mpirun to run this in Linux. For example you can do the following within the directory of the executable,

    mpirun -np 4 mono ./Program.exe
  

  which will produce,

    Rank 0 of 4 running on i81
    Rank 2 of 4 running on i81
    Rank 1 of 4 running on i81
    Rank 3 of 4 running on i81
  

  where i81 is one of the compute nodes in FutureGrid cluster.
  You may also use other advance options with mpirun to determine process mapping and binding. Note. the syntax for such controlling is different from latest versions of OpenMPI. Therefore, it's a good idea to look at different options from mpirun --help. For example you may be interested in specifying the following options,

    hostfile=<path-to-hostfile-listing-available-computing-nodes>
    ppn=<number-of-processes-per-node>
    cpp=<number-of-cpus-to-allocate-for-a-process>
  
    mpirun --display-map --mca btl ^tcp --hostfile $hostfile --bind-to-core --bysocket --npernode $ppn --cpus-per-proc $cpp -np $(($nodes*$ppn)) ...
  

  where, --display-map will print how processes are bind to processing units and --mca btl ^tcp forces to turn off tcp