Running MPI.NET Applications with Mono in Ubuntu

Sometime back I played around Mono to get some of our parallel applications running on Ubuntu. These applications were C# based and used MPI.NET.

The following blog post is a great starting point with all the details you'll need. So I'll skip the steps, except to point out couple of caveats you need to consider.

http://blog.biophysengr.net/2011/11/compiling-mpinet-under-ubuntu-oneiric.html

1. automake versions above 1.9 will give you an error when building MPI.NET. May be you can change the make script to work with them, but I found it easy to just install automake1.9 to solve it.
2. You'll need to add /usr/local/lib to your LD_LIBRARY_PATH. Essentially what you need to do is add this path to /etc/ld.so.conf and run ldconfig as root. See documenation from Mono on this at http://www.mono-project.com/DllNotFoundException
3. Make sure to do chmod +x to your dlls

Repair Mr.Coffee (IDS77) Thermal Fuse

If you haven't bought this product yet, then STOP don't buy it!!

If you, however, have bought it and broke it in the first run then continue.

Mine simply stopped working right on the day I bought it, in fact this is the second Mr. Coffee grinders I bought that day. May be I was trying to grind too much, but as a consumer device I'd expect it to "auto shut off" if it's too hot, rather burn itself.

The good news is, it only burns a thermal fuse, which fairly easy to replace if you get under the hood. Once you remove the grinder cup you can see the following,

Just remove the three plastic hole cover knobs, which you can simply pull out using tweezers or by lifting one side with a sharp point like that of a knife. Once removed these knobs look as shown below.

Get a small flat head screw driver and remove the three screws. Then you can simply take out the motor compartment. You'll need to pop the button panel to find a bolt holding the circuit board. You'll be able to figure out hopefully.

Once the motor is out you can see the thermal fuse wrapped inside the yellow tape covering the motor winding as shown. I cut the pins of the fuse and it's shown left to the motor. 

Now it's time to find a replacement. The original fuse that's there in this one comes from China and here's a link I could find on it. A close up picture is given below.

I found several options in ebay, but either they had low ampere rating or too high functioning temperature. Also, it'd take more than a week to arrive. In the end, I decided to go with an alternative one from RadioShack, which in fact is cheaper (~$1.40) than options from ebay.

If you need some instructions on how to solder these look at this. This one has a high ampere rating, but from what I read having a higher ampere rating than the one used does no harm. It's the temperature that's important.

Once soldered, the unit is alive again. This time I will not grind coffee continuously though :)

Home Made Meditation Benches

My friend, Lahiru, suggested  to make few meditation benches as a donation to Indiana Buddhist Temple few weeks back. I had no idea what a meditation bench was, but luckily he had a basic one with him. After few modifications to the design, here’s what we ended up with.

If you like to make one at home too, here are the details.

• Materials
• Wood – 1 x 8 (i.e. 3/4” x 7+1/4”) pine boards http://www.menards.com/main/building-materials/lumber-boards/boards/select-boards/1-x-8-x-8-select-pine-board/p-1934090-c-13118.htm
• Hinges – 3” x 3” door hinges, 1/4” round corner http://www.menards.com/main/tools-hardware/builders-hardware/door-window/residential-hinges/3x3-door-hinge/p-1475757-c-9687.htm
• Used 1/2” #10 Phillips flat head screws http://www.menards.com/main/building-materials/drywall/screws/wood/10-x-1-2-phillips-flat-wood-100pcs-pkg/p-1638963-c-8940.htm
• Leg sliders – Slipstick 3/4” Slider Foot http://www.amazon.com/Slipstick-CB190-4-Inch-Slider-16-Pack/dp/B000G9WK44/ref=sr_1_2?ie=UTF8&id=1365832735&sr=8-2&keywords=slipstick+slider
• Felt pads (not shown in picture) - http://www.amazon.com/Waxman-Self-Stick-Round-Brown-8-Inch/dp/B000VYN7CS/ref=sr_1_4?s=hi&ie=UTF8&qid=1365832847&sr=1-4&keywords=felt+pads
• De-waxed Shellac – Bulls Eye Seal Coat http://www.menards.com/main/interior-wood-care/specialty-wood-finishes/bulls-eye-seal-coat-universal-sealer/p-1963335-c-13129.htm
• Steps
• Cutting wood – a hand saw is enough actually, but I used a jig saw to make it bit faster and to cut the semi circles on legs. Dimensions are given below

For legs, cut from the dashed lines, which are 1/2” from each edge. The 5.25 and 6.75 are lengths along the outside edges though. Alternatively, you could determine the angle from these two values and then calculate the the lengths for the actual cut piece.	The jig I made to draw semi circles and flush legs to the size. About the semi circle, I just used one of my wife’s planting pots to draw it. So I didn’t measure its diameter. You can use whatever shape or size you like.

• Fixing hinges – I used a router to make the mortises so that legs would close flush with the main board.You could omit this if this is not a concern.
• NOTE. I have flipped one side of the hinge (by removing its pin) to get them fixed as shown above. The reason was that the hinges I used won’t close flushed. See diagrams below.
• I didn’t like the protruding wedge shape (right most image above), hence the flipping (see here to know how to remove hinge pin) and mortising.
• Now, with some custom mortise routing I could get the legs to close flushed (above right). I made few jigs to make it easy to batch these as below.

Mortise on the main board (ignore the writing), it’s the same for legs except the groove for hinge pin is on the edge rather in the middle and only half of the groove is necessary.	Simple jig made from 1/2” MDF. The dimensions for the hole depends on the flush trim router bit you are going to use. I used a 1/2” bit that came with this mortise kit. Roughly mine had about 1/16” added to actual hinge dimensions
I had to use this 1/8” hardboard mask over the jig to get the depth difference in the mortise for groove and the hinge plate. So the process was to route entire shape WITH mask and then route the groove only WITHOUT mask. A plunge-based router would have eliminated this need.	Cross section of the mortise

• Finishing – 120 grit sanding followed by de-waxed shellac with 220 sanding between coats.

I think this covers most of the stuff you need to complete this project and any questions are welcome in comments. Happy bench making !!

Notes on Windows PowerShell

• Is bit slow and creepy ;) but better than command line
• Running as administrator helps you overcome many of the access denied situations
• PowerShell script files are simple text files with .ps1 extension
  • To run a script file,
    • Open a PowerShell instance (preferably as administrator) 
    • Type the path to .ps1 file and hit enter
  • If you get an error due to execution-policy is restricted, try

    set-executionpolicy remotesigned 

  • The above will ask for your permission. To avoid that and force use

     set-executionpolicy remotesigned -force

  • If you want this to be done across a Windows HPC cluster using clusrun command

    $nodes="node1,node2,..,noden"
    clusrun /nodes:$nodes powershell set-executionpolicy remotesigned -force

• To start a new process through a script - http://technet.microsoft.com/en-us/library/hh849848.aspx

  start-process command "options and args"

  
  
  • If you want it to be on the same window and wait for it to complete use

    start-process command "options and args" -NoNewWindow –Wait

• Invoke a command on a remote machine - http://technet.microsoft.com/en-us/library/hh849719.aspx

Invoke-Command -ComputerName $node -ScriptBlock {Start-Service MyService}


Handling variables


• Local variables

$x=10
$path="C:\users\dd"



• Command line parameters

$param0=$args[0]
$param1=$args[1]



• Environment (User/System) variables

$javaBin=$env:JAVA_HOME + "\bin"



• A better alternative – number 2 indicates system variable. Use number 1 for user variable

$javaHome=[Environment]::GetEnvironmentVariable("JAVA_HOME",2)



• Similarly to set a value – again number 2 is for system level.

[Environment]::SetEnvironmentVariable("Path","$tmp",2)

• String operations (few)


• Concatenation

$name="John"
$x="hello " + $name + "! How are you ?"



• Substring

$name.Substring(1) // "ohn"
$name.Substring(1,2) // "oh"



• StartsWith

$name.StartsWith("hello") // true



• IndexOf

$name.IndexOf("ohn") // 1



• Length

$name.Length // 4



• Join paths

$javaHome=$env:JAVA_HOME // e.g. "C:\Program Files\Java\jdk1.7.0_10"
$javaBinString=join-path -path $javaHome "bin" // e.g. "C:\Program Files\Java\jdk1.7.0_10\bin"



• Loops – a good tutorial at http://www.powershellpro.com/powershell-tutorial-introduction/logic-using-loops/


• For loop

for($i=1;$i -le 10; $i++) {
  //body
}

• -le is <=
• -lt is <
• -gt is >
• -eq is == in usual programming language notation


• Foreach loop

$jars=ls ($env:TWISTER_HOME + "\lib") *.jar
foreach ($jar in $jars) {
 $jar=$jar.DirectoryName + "\" +  $jar.Name
 $cp=$jar+";"+$cp
}

• .NET runtime directory

$([System.Runtime.InteropServices.RuntimeEnvironment]::GetRuntimeDirectory())

This is pretty much I found useful for my work. Feel free to suggest any.

Study of Biological Sequence Structure: Clustering and Visualization

My presentation in our SALSA HPC weekly meeting.
http://salsahpc.blogspot.com/2013/02/study-of-biological-sequence-structure.html

K-Means Clustering with Chapel

K-Means clustering is simple to comprehend, yet is a good candidate to benefit from parallelization. The idea is to partition N data points into K clusters by assigning points to the nearest cluster center. The program starts with a random K number of centers. These centers are then refined iteratively by taking the mean of nearest points to each center. The refinement continues until the change in centers falls below a predefined threshold or until a fixed number of iterations have completed.

Full working code is at https://www.dropbox.com/s/ikfktb3sswmc8vy/kmeans.chpl

• Declaration of points, dimension of a point, iterations, and number of clusters. Note. These are configurable at start time by passing them as command line arguments. For example passing --numPoints 50000 will set the number of points to 50000 instead of 2000.

     config var numDim: int = 2,
                numClusters: int = 4,
                numPoints: int  = 2000,
                numIter: int = 50,
                threshold: real = 0.0001;

• Definitions of domains. The {0 .. #N} notation indicates the integer range starting at zero and counting up to N number of values, i.e. 0, 1, 2 ... N-1

     const PointsSpace = {0..#numPoints,0..#numDim};
     const ClusterSpace = {0..#numClusters,0..#numDim};
     const ClusterNumSpace = {0..#numClusters};

• Block distribute points along zeroth dimension across locales. The array Locales and the numLocales are made available to programs by the Chapel runtime and we have reshaped the locales into a two dimensional array of numLocales x 1 size. This ensures components of each point stay in the same locale when blocked. 

     var blockedLocaleView = {0..#numLocales,1..1};
     var blockedLocales: [blockedLocaleView] locale = reshape(Locales, blockedLocaleView);
     const BlockedPointsSpace = PointsSpace 
       dmapped  Block(boundingBox=PointsSpace, targetLocales=blockedLocales);
     var points: [BlockedPointsSpace] real;

• Arrays to hold current centers. Chapel’s rich domain operations make it possible to assign a subset of points as current centers by specifying a range as shown in second line. 

     var currCenters: [ClusterSpace] real;
     currCenters = points[ClusterSpace];  

• Replicated arrays to keep local centers and updates. Chapel allows transparent access to array elements on remote locales; however, performance may suffer when data is transferred over network. Therefore, it is beneficial to use local arrays when operating on points in a particular locale. Chapel provides a convenient distribution to achieve this, ReplicatedDist, which creates a copy of the array in each locale. The array reference resolves to the local copy in the particular locale where the referring code runs. Also note the use of atomic real and integer variables, which is done as a workaround to non-implemented atomic blocks. 

     const ReplClusterSpace = ClusterSpace dmapped ReplicatedDist();
     var localCurrCenters: [ReplClusterSpace] real;
     // using atomic primitives as a work around to not implemented atomic blocks
     var localCenterUpdates: [ReplClusterSpace] atomic real;
     const ReplClusterNumSpace = ClusterNumSpace dmapped ReplicatedDist();
     var localCenterPCounts: [ReplClusterNumSpace] atomic int;

• The next steps happen iteratively while refining centers.
• We start by resetting local arrays as follows. The first line copies the current centers to localCurrCenters array in each locale. The ReplicatedDist in Chapel guarantees the array assignment in the first line happens in each locale. The next two lines initialize the two local arrays, i.e. localCenterUpdates and localCenterPCounts, to zero. Again, the distribution guarantees the two forall loops happen for each local copy of the arrays. These three statements are run in parallel by wrapping inside a cobegin clause.          

          cobegin {
            localCurrCenters = currCenters;
            forall lcu in localCenterUpdates do lcu.write(0.0);
            forall lcpc in localCenterPCounts do lcpc.write(0);
          }

• Next is to compare the distance for each point against all cluster centers and decide the cluster it belongs to.  Note the shifting of locales using the on clause in line 2, which in turn guarantee the access of current centers, center updates, and center point counts arrays local to the particular locale. We have placed the atomic construct to show where the updates should be done atomically, but it is not implemented in Chapel yet. However, the use of atomic variables, overcomes this and guarantees proper atomic updates. Also note if the number of clusters was large, the for loop could be changed to the parallel forall version with slight modification to the code. Moreover, if parallel task creation was unnecessarily expensive one could easily change the code to use serial execution.

          forall p in {0..#numPoints} {
            on points[p,0] {
              var closestC: int = -1;
              var closestDist: real = MaxDist;
              for c in {0..#numClusters} { // possibility to parallelize
                var dist: atomic real;
                dist.write(0.0);
                forall d in {0..#numDim} {
                  var tmp = points[p,d] - localCurrCenters[c,d];
                  dist.add(tmp*tmp);
                }
                if (dist.read() < closestDist) {
                  closestDist = dist.read();
                  closestC = c;
                }
              }
              forall d in {0..#numDim} {
                atomic { // here's where we need atomic
                  localCenterUpdates[closestC,d].add(points[p,d]);
                }
              }
              localCenterPCounts[closestC].add(1);
            }
          }

• Then we collate centers in each locale. Note the use of nested forall and cobegin constrcuts. Also, the tmpLCU and tmpLCPC arrays are not distributed, yet Chapel gives seamless access to their elements even when the referring code runs on remote locales.

          var tmpLCU: [ClusterSpace] atomic real;
          forall tlcu in tmpLCU do tlcu.write(0.0);
          var tmpLCPC: [0..#numClusters] atomic int;
          forall tlcpc in tmpLCPC do tlcpc.write(0);

          forall loc in Locales {
            on loc do {
              cobegin {
                forall (i,j) in ClusterSpace {
                  tmpLCU[i,j].add(localCenterUpdates[i,j].read());
                }
                forall i in {0..#numClusters} {
                  tmpLCPC[i].add(localCenterPCounts[i].read());

                }
              }
            }
          }

• Finally, we test for convergence. If it has converged or the number of iterations has exceeded the given limit we will stop and will print results.

          var b: atomic bool;
          b.write(true);
          forall (i,j) in ClusterSpace {
            var center: real = tmpLCU[i,j].read()/tmpLCPC[i].read();
            if (abs(center - currCenters[i,j]) > threshold){
              b.write(false);
            }
            currCenters[i,j] = center;
          }
          converged = b.read();

P.S. I thought I should mention here that this code may not be the most optimal due to the fact that for each point it'll do a locale shift. Ideally, I'd like to implement this in a fashion as "for each locale do for each point", but this requires you knowing the points assigned to a particular locale. Currently, Chapel doesn't have a readily available method to retrieve this info. However, Chapel developers suggested some workaround to my mail on Identifying Local Indices of Distributed Arrays

Automate Phylogenetic Tree Coloring: Dendroscope

Dendroscope (http://ab.inf.uni-tuebingen.de/software/dendroscope/welcome.html) is a neat tool that we (SALSA group in IU) have been using in our research work related to phylogenetic trees. Recently we encountered a scenario where leaf nodes need to be colored according to a corresponding cluster number. We already had generated the mapping from leaf node labels to RGB color strings. To complicate it further, we wanted to color the “lowest single ancestor” (LSA) induced sub trees for each cluster in the same color.

Hard Labor

Open tree file using Dendroscope –> select leaf nodes for a particular cluster –> click “Select –> Advanced Selection –> LSA Induced Network” –> Ctrl+J –> check “Label Color” and “Line  Color” –> pick the desired color –> hit “close” –> click somewhere to unselect nodes/edges
Repeat this (except opening file) for each cluster.

Hard Labor with Pain Killer

You can avoid the step in bold above by having a text file with label names for the particular cluster. Then after opening the tree file you can hit Ctrl+f and specify your text file there (using the open folder icon in that). This will go through each line of the text file and select the labeled leaf nodes. This avoids having to click numerous times just to select :). Still you have to go through the pain of selecting the LSA sub tree and modifying colors.

No Pain, YES! gain

The clean way is to automate this process and Dendroscope facilitates this through its command line. Here’s what you need to generate using a simple script written by you.
open file=<path-to-your-tree-file>;select all;set edgewidth=2;set color=0 0 0;set labelcolor=0 0 0;deselect all;
find searchtext=abc;find searchtext=def
select LSA induced network;set color=r g b;set labelcolor=r g b;deselect all;
Some explanation please …

Line 1 will simply open your tree file and set all the edges to a width of 2 and everything to black
For each label in a particular cluster you need to add find searchtext=<label>; as in Line 2 (remember to separate with semi-colon)
Line 3 will select the LSA sub tree for the labels in Line 2 and line and label color to whatever color specified by r g b values
You need to include corresponding Line 2 and 3 for each cluster.
Once you generate this file simply open Dendroscope and go to “Window –> Command Input”. In the pop-up command window type source file=<path-to-your-generated-file> and hit “Apply”. That’s it enjoy!!

A good reference is available in Dendroscope’s manual at http://ab.inf.uni-tuebingen.de/data/software/dendroscope3/download/manual.pdf

How to Compute The Sum of Squares for First N Integers

We need to find,
	answer
How to get that?
Let’s consider the expansion of
Therefore,
Now let’s substitute values for
	when
	when
…
	when
	when
if you sum them from top to bottom
Taking the sum of arithmetic series,
Therefore,

Special thanks to my friend Jayampathi Rathnayake (http://www.math.indiana.edu/people/profile.phtml?id=jratnaya) for his insight on the approach.

Another good resource http://www.trans4mind.com/personal_development/mathematics/series/sumNaturalSquares.htm