## Sunday, July 15, 2012

### Julia programming language

Note: working with Julia commit 2013-07-26 (0.2.0-prerelease+2820)

Julia is a new high-level, high-performance dynamic programming language for technical computing.
High-performance assures fast execution, whereas dynamic languages enable fast development.

Both used to be contradictionary features of programming languages, since dynamic basically means no compilation before execution, and that means: More work at run-time. However, since Just-In-Time-Compilers got better and better, now there is a way to have dynamic and high perfomance languages. Julia is one of those languages. They use the LLVM as JIT-compiler - which I think is a pretty neat idea.

Installation is moderately easy as is described here . There are some precompiled packages here, but since Julia is pretty new language, it is best to do a git clone . For that, you might want to update your GCC, since I was not able to compile it with 4.1.2 which comes standard with OSX.

### Julia sets

So somehow I could not resist in writing a program calculating a Julia set in Julia. Since there was already the mandelbrot example here, it was easy to change it to calculate Julia sets:

julia.jl
# the julia iteration function
function julia(z, maxiter::Int64)
c=-0.8+0.156im
for n = 1:maxiter
if abs(z) > 2
return n-1
end
z = z^2 + c
end
return maxiter
end


Since Julia has built-in support for complex numbers, we can now calculate, for example julia(0.5+0.2im, 256)

### Sequential Calculation

To calculate a whole picture, we need some for loops like this:
calcJulia.jl
# load image support
include("myimage.jl")

include("julia.jl")

# create a 1000x500 Array for our picture
h = 500
w = 1000
m = Array(Int64, h, w)

# time measurements
print("starting...\n")
tStart=time()

# for every pixel
for y=1:h, x=1:w
# translate numbers [1:w, 1:h] -> -2:2 + -1:1 im
c = complex((x-w/2)/(h/2), (y-h/2)/(h/2))
# call our julia function
m[y,x] = julia(c, 256)
end

tStop = time()

# write the ppm-file
myppmwrite(m, "julia.ppm")

print("done. took ", tStop-tStart, " seconds\n");


Unfortuately, there is no image.jl anymore shipped with julia anymore, so here is a very easy version of it:
myimage.jl
function myppmwrite(img, file::String)
s = open(file, "w")
write(s, "P6\n")
n, m = size(img)
write(s, "$m$n 255\n")
for i=1:n, j=1:m
p = img[i,j]
write(s, uint8(p))
write(s, uint8(p))
write(s, uint8(p))
end
close(s)
end


Together with the julia.jl file, you can execute above script with
julia calcJulia.jl

There is an even easier way to calculate the array m:
m = [ julia(complex(r,i)) for i=-1.0:.01:1.0, r=-2.0:.01:0.5 ];
(taken from extras/image.jl). However, above approach yields more insight on the parallel algorithm.

### Parallel Calculation

Parallel programming is not as easy as in, for example OpenMP.  We are creating a distributed array here (darray) and initialize it with the function parJuliaInit, which has to calculate its local part of the array. Because every processor needs to know the init function parJuliaInit and julia, we need to use the @everywhere command for the load and the function definition (@everywhere is not explained in the docs yet):

parJulia.jl
include("myimage.jl")
# we need julia.jl everywhere
@everywhere include("julia.jl")

@everywhere w=1000
@everywhere h=500

# the parallel init function, needed everywhere
@everywhere function parJuliaInit(I)
# create our local patch
d=(size(I[1], 1), size(I[2], 1))
m = Array(Int, d)

# we need to shift our image
xmin=I[2][1]
ymin=I[1][1]

for x=I[2], y=I[1]
c = complex((x-w/2)/(h/2), (y-h/2)/(h/2))
m[y-ymin+1, x-xmin+1] = julia(c, 256)
end
return m
end

print("starting...\n")
tStart = time()

# create a distributed array, initialize with julia values
Dm = DArray(parJuliaInit, (h, w))

tStop = time()

# convert into a local array
m = convert(Array, Dm)

# write the picture
myppmwrite(m, "julia.ppm")

# print out time
print("done. took ", tStop-tStart, " seconds\n");



### Results

You can run that code with
julia -p 4 parJulia.jl
Where you replace 4 with your number of processors.

Here's a comparison of different processor sizes and algorithms for different picture sizes calculated on a MacPro with 2x 2.26 Ghz Quad-Core Xeon processors (= 8 processors in total):

code #processors time 500x1000 time 2000x4000
calcJulia.jl 1 2.20 s 34.50 s
parJulia.jl 1 3.16 s 46.26 s
parJulia.jl 2 1.83 s 23.64 s
parJulia.jl 4 0.92 s 7.28 s
parJulia.jl 5 0.83 s 5.27 s
parJulia.jl 8 0.85 s 2.97 s
parJulia.jl 10 0.81 s 2.50 s

The table shows the difficulties in parallel computing: Even if the algorithm scales perfectly in theory, the execution on N processors is not always N times faster. This is because of not all parts of your implementation and your hardware might scale perfectly, especially not for all problem sizes (very large or very small). The drastically difference of access rates for the L1/L2/RAM memory can even lead to results like the parallel calculation of the 2000x4000 image, which is 15 times faster on 8 processors than on one.

### Conclusion

Julia is a nice language with a MATLAB-style syntax which could have a big influence on scientific computing. Many applications are operating on the memory bandwith limit, or the communication bandwith limit, so with a JIT compiler, those would be just fine.

Note: Syntax highlighting is done with Syntax Highlighter and this little beta JavaScript configuration file for the Julia language.

Update 1:
Syntax Highlighting should now work.
Update 2:
Fixed sequential calculation.
Update 3:
Fixed parallel calculation. load() is not @everywhere by default.
Update 4:
Now working with Julia commit 2013-07-26, which changed DArray syntax and require.