start:development:scripting_languages:julia [MogonWiki]

This page is read only. You can view the source, but not change it. Ask your administrator if you think this is wrong.
====== Julia Programming Language  ====== 

<callout type="info" icon="true" title="Work In Progress">
<progress>
<bar value="33" type="danger" striped="true" animate="true"></bar>
<bar value="33" type="warning" striped="true" animate="true"></bar>
<bar value="10" type="success" striped="true" animate="true"></bar>
</progress>
</callout>


[[https://julialang.org/|Julia]] is a high-performance, high-level, dynamic programming language.

<grid>
<col lg="1" md="1" sm="1" xs="1"></col>
<col lg="10" md="10" sm="10" xs="10">
<blockquote>
Distinctive aspects of Julia's design include a type system with parametric polymorphism in a dynamic programming language; with multiple dispatch as its core programming paradigm. Julia supports concurrent, (composable) parallel and distributed computing (with or without using MPI and/or the built-in corresponding to "OpenMP-style" threads), and direct calling of C and Fortran libraries without glue code. Julia uses a just-in-time (JIT) compiler that is referred to as "just-ahead-of-time" (JAOT) in the Julia community, as Julia compiles all code (by default) to machine code before running it.

<cite>[[https://en.wikipedia.org/wiki/Julia_(programming_language)|Wikipedia]]</cite>
</blockquote>
</col>
<col lg="1" md="1" sm="1" xs="1"></col>

</grid>



==== Julia Modules ====

We currently offer the following Julia versions:

<code bash>
JGU HPC Modules
------------------- /cluster/easybuild/broadwell/modules/all -------------------
   lang/Julia/1.5.3-linux-x86_64    lang/Julia/1.6.3-linux-x86_64
   lang/Julia/1.6.0-linux-x86_64    lang/Julia/1.7.0-linux-x86_64 (D)
</code>

One can use a specific version with the following command:

<code bash>
module load lang/Julia/1.6.0-linux-x86_64
</code>

or the current default module with:

<code bash>
module load lang/Julia
</code>

====== Pkg - Julia's Package Manager ======

[[https://pkgdocs.julialang.org/v1/|Pkg]] is Julia's built-in package manager that performs tasks such as installing, updating and removing packages.

==== Pkg - Usage ====

Initially, you need to load a Julia Module on a MOGON service-node, e.g. with:

<code bash>
module load lang/Julia
</code>

For package installation, we will use the REPL (read-eval-print loop) that comes built-in to the ''julia'' executable. Start Julia by using:

<code bash>
julia
</code>

from the command line. Now start the Pkg REPL (Pkg also comes with a REPL) by pressing '']'', uppon entering the Pkg REPL, the command line prompt should like:

<code bash>
(1.6) pkg>
</code>

We will use the package **Dates** ([[https://docs.julialang.org/en/v1/stdlib/Dates/|Dates]]) to illustrate what the general procedure for installing packages is:

<code bash>
(1.6) pkg> add Dates
</code>

you should get an output similar to:

<code julia>
(v1.6) pkg> add Dates
   Resolving package versions...
    Updating `/gpfs/fs1/home/<username>/.julia/environments/v1.6/Project.toml`
  [ade2ca70] + Dates
  No Changes to `/gpfs/fs1/home/<username>/.julia/environments/v1.6/Manifest.toml`
</code>

Let's check the successful installation: First we display the status of the packages in our standard project:

<code julia>
(v1.6) pkg> status
</code>

Depending on which packages you have already installed, you should get an output similar to the following:

<code julia>
(v1.6) pkg> status
      Status `/gpfs/fs1/home/<username>/.julia/environments/v1.6/Project.toml`
  [6e4b80f9] BenchmarkTools v0.7.0
  [052768ef] CUDA v3.0.3
  [7a1cc6ca] FFTW v1.3.2
  [da04e1cc] MPI v0.17.2
  [91a5bcdd] Plots v1.11.2
  [d330b81b] PyPlot v2.9.0
  [ade2ca70] Dates
  [de0858da] Printf
  [9a3f8284] Random
</code>

This indicates that ''Dates'' has been successfully installed, but let's check and confirm this with ''Pkg''

<code julia>
(v1.6) pkg> test Dates
     Testing Dates
      Status `/tmp/jl_cb9pkw/Project.toml`
  [ade2ca70] Dates `@stdlib/Dates`
  [b77e0a4c] InteractiveUtils `@stdlib/InteractiveUtils`
  [de0858da] Printf `@stdlib/Printf`
  [8dfed614] Test `@stdlib/Test`
      Status `/tmp/jl_cb9pkw/Manifest.toml`
  [2a0f44e3] Base64 `@stdlib/Base64`
  [ade2ca70] Dates `@stdlib/Dates`
  [b77e0a4c] InteractiveUtils `@stdlib/InteractiveUtils`
  [56ddb016] Logging `@stdlib/Logging`
  [d6f4376e] Markdown `@stdlib/Markdown`
  [de0858da] Printf `@stdlib/Printf`
  [9a3f8284] Random `@stdlib/Random`
  [9e88b42a] Serialization `@stdlib/Serialization`
  [8dfed614] Test `@stdlib/Test`
  [4ec0a83e] Unicode `@stdlib/Unicode`
     Testing Running tests...  
[ ... ]
Test Summary:               | Pass  Total
Conversions to/from numbers |   19     19
     Testing Dates tests passed
</code>

The installation was sucessful and ''Dates'' can now be included in ''.jl'' file via the following line:

<code julia>
using Dates
</code>

Yoa can now add any packages from the Standard Library to Julia, but please note the following:

<callout type="warning" title="Package Dependencies " icon="true">
Some Julia packages require you to load pre-requisite dependencies as modules before you can add the via ''Pkg.add''!
</callout>

This is also illustrated again by the following examples [[start:development:scripting_languages:julia#pkg_-_cuda|CUDA]] and [[start:development:scripting_languages:julia#pkg_-_plots|Plots]].

=== Pkg - Commands ===
In the Pkg REPL you have the following commands available to manage packages:

^ Command ^ Result ^ Comment^
|''add'' | Adds a Package | It is possible to ad multiple packages in one command ''add A B C'' |
|''test''| Run Tests for a Package | |
|''build''| Explicitly run the build step for a Package | ''build'' is automatically run when a package is first installed with ''add''|
|''rm'' | Removes Package | Removes only that package in the project. Tto remove a package that only exists as a dependency use ''rm --manifest DepPackage''. Note that this will remove all packages that depend on ''DepPackage''. |
|''up'' | Update Package and //all// dependencies | The Flag ''--minor'' updates only the minor version of packages reducingthe risk to break projects.|
|''status''| Print out the status of the project/manifest.| |


Installing packages using ''Pkg'' adds them to the default project at ''~/.julia/environments/<JuliaVersion>''. However, it is also possible to work with Environments in Julia and [[https://pkgdocs.julialang.org/v1/environments/#Creating-your-own-projects|create own projects]], [[https://pkgdocs.julialang.org/v1/environments/#Precompiling-a-project|precompiling a project]] and even [[https://pkgdocs.julialang.org/v1/environments/#Using-someone-else's-project|using someone else's project]]

==== Pkg - CUDA ====

Julia's [[https://juliagpu.gitlab.io/CUDA.jl/|CUDA.jl]] package is the main entrypoint for programming on NVIDIA GPUs using CUDA. The Julia CUDA stack requires a functional CUDA-setup, which includes the NVIDIA Driver and the corresponding CUDA toolkit. These are either available as module or already integrated into the MOGON GPU nodes. To be able to use ''CUDA.jl'' with Julia, you only need to proceed as follows.

Log in to MOGON and load the following modules on a service-node first:

<code bash>
module load system/CUDA/11.0.3-GCC-9.3.0
module load lang/Julia
</code>

Now start Julia with the following command

<code bash>
julia
</code>

and then change to the Pkg REPL with '']'', the command line prompt should now look like:

<code julia>
(v1.6) pkg>
</code>

We are now ready to add CUDA via

<code julia>
(v1.6) pkg> add CUDA
</code>

You are now ready to use ''CUDA.jl'' with Julia on MOGON. Take a look at our GPU section below for easy approach to ''CUDA.jl'' on MOGON.

==== Pkg - Plots ====

[[https://docs.juliaplots.org/latest/|Plots.jl]] is a visualization interface and toolset for powerful visualizations in Julia. Here we explain how to add the ''Plots.jl'' package to Julia and set up the backend ''PyPlot''.
Log in to MOGON and load the following modules on a service-node first:

<code bash>
module load lang/SciPy-bundle/2020.03-foss-2020a-Python-3.8.2
module load vis/matplotlib/3.2.1-foss-2020a-Python-3.8.2
module load lang/Julia
</code>

Open Julia by executing the following command after the modules have been successfully loaded

<code bash>
julia
</code>

now enter the Pkg REPL by pressing '']'', the command line prompt should look like:
<code julia>
(v1.6) pkg>
</code>
First, the actual packages are added and then the backend is configured. Install ''Plots.jl'' with:

<code julia>
(v1.6) pkg> add Plots
</code>

now set the backend to ''pyplot'' with:


<code julia>
(v1.6) pkg> add PyPlot
</code>
Afterwards, test the successful installation with:
<code julia>
(v1.6) pkg> test Plots
</code>

and

<code julia>
(v1.6) pkg> test PyPlot
</code>

==== The Julia Standard Library ====

An overview of available packages for Julia can be found in the [[https://docs.julialang.org/en/v1/|Julia Documentation]] and [[https://github.com/JuliaLang/julia/tree/master/stdlib|JuliaLang Github Repo]]. The most noteworthy packages are (aka. you will probably install them at some point):

^ Package ^ Comment ^
| Random | | 
| CUDA | |
| LinearAlgebra | |
| Printf| |
| MPI | |
| BenchmarkTools||
| FFTW | |


====== Submitting a Serial Julia Job ======

<file julia hello_mogon.jl>
println("Hello MOGON!")
</file>

<file bash serial_julia_job.slurm>
#!/bin/bash
#SBATCH --partition=smp
#SBATCH --account=<YourHPC-Account>
#SBATCH --time=0-00:01:00
#SBATCH --mem=512 #0.5GB
#SBATCH --ntasks=1
#SBATCH --job-name=julia_serial_example
#SBATCH --output=%x_%j.out
#SBATCH --error=%x_%j.err
 
module purge
module load lang/Julia
 
julia hello_mogon.jl
</file>

<code bash>
cat julia_serial_example_*.out
</code>

<code bash>
Hello MOGON!
</code>
====== Submitting a Parallel Julia Job ======

Julia offers [[https://docs.julialang.org/en/v1/manual/parallel-computing/|two main possibilities]] for parallel computing: A **multi-threading** based parallelism, which basically is a shared memory parallelism and **distributed Processing**, which parallelizes code across different Julia processes.

===== Multi-Threading =====

<blockquote>
The number of execution threads is controlled either by using the ''-t/--threads'' command line argument or by using the ''JULIA_NUM_THREADS'' environment variable. When both are specified, then ''-t/--threads'' takes precedence.

<cite>Julia Documentation, [[https://docs.julialang.org/en/v1/manual/multi-threading/#man-multithreading|Multi-Threading]]</cite>
</blockquote>

Therefore, to start Julia with four threads, you must execute the following command:

<code bash>
julia --threads 4
</code>

But let's explore the basics of Julia's multi-threading capabilities with an example:

<file julia hello_mogon_smp.jl>

Threads.@threads for i=1:20
    println(Hello MOGON! The Number of iteration is $i from Thread $(Threads.threadid())/$(Threads.nthreads()))
end    
</file>

<file bash julia_smp_job.slurm>
#!/bin/bash
#SBATCH --partition=smp
#SBATCH --account=<Your-ZDV-Account>
#SBATCH --time=0-00:02:00
#SBATCH --mem-per-cpu=1024 #1GB
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=6
#SBATCH --job-name=smp_julia
#SBATCH --output=%x_%j.out
#SBATCH --error=%x_%j.err

module purge
module load lang/Julia

julia --threads $SLURM_CPUS_PER_TASK hello_mogon_smp.jl
</file>

Once the job is finished, you can display the result with the following command:
<code bash>
cat smp_julia_*
</code>

The output should be similar to the following lines:

<code bash>
Hello MOGON! The Number of iteration is 5 from Thread 2/6
Hello MOGON! The Number of iteration is 9 from Thread 3/6
Hello MOGON! The Number of iteration is 15 from Thread 5/6
Hello MOGON! The Number of iteration is 10 from Thread 3/6
Hello MOGON! The Number of iteration is 11 from Thread 3/6
Hello MOGON! The Number of iteration is 12 from Thread 4/6
Hello MOGON! The Number of iteration is 16 from Thread 5/6
Hello MOGON! The Number of iteration is 17 from Thread 5/6
Hello MOGON! The Number of iteration is 6 from Thread 2/6
Hello MOGON! The Number of iteration is 13 from Thread 4/6
Hello MOGON! The Number of iteration is 7 from Thread 2/6
Hello MOGON! The Number of iteration is 14 from Thread 4/6
Hello MOGON! The Number of iteration is 1 from Thread 1/6
Hello MOGON! The Number of iteration is 8 from Thread 2/6
Hello MOGON! The Number of iteration is 2 from Thread 1/6
Hello MOGON! The Number of iteration is 3 from Thread 1/6
Hello MOGON! The Number of iteration is 18 from Thread 6/6
Hello MOGON! The Number of iteration is 4 from Thread 1/6
Hello MOGON! The Number of iteration is 19 from Thread 6/6
Hello MOGON! The Number of iteration is 20 from Thread 6/6
</code>

===== Distributed Processing =====

<blockquote>
Starting with ''julia -p n'' provides ''n'' worker processes on the local machine. Generally it makes sense for ''n'' to equal the number of CPU threads (logical cores) on the machine. Note that the ''-p'' argument implicitly loads module Distributed

<cite>Julia Documentation, [[https://docs.julialang.org/en/v1/manual/distributed-computing/|Multi-processing and Distributed Computing]]</cite>
</blockquote>


<file julia distributed_julia_example.jl>
@everywhere begin
    using LinearAlgebra
    a = zeros(200,200);
end

slurm_cores = parse(Int, ENV["SLURM_CPUS_PER_TASK"])
slurm_tasks = parse(Int, ENV["SLURM_NTASKS"])

println("Number of requested Slurm Tasks is: ", slurm_tasks)
println("Number of requested Slurm CPUs per Task is: ", slurm_cores)

println("Number of available Julia Processes: ", nprocs())
println("Number of available Julia Worker Processes: ", nworkers())

calctime = @elapsed @sync @distributed for i=1:200
                a[i] = maximum(abs.(eigvals(rand(500,500))))
           end
           
println("With ", slurm_cores, " CPUs per Task the calculation took ", calctime, " seconds.")                      
    
</file>

<file bash julia_distributed_job.slurm>
#SBATCH --partition=smp
#SBATCH --account=<Your-HPC-Account>
#SBATCH --time=0-00:03:00
#SBATCH --mem-per-cpu=4096 #4GB
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=2
#SBATCH --job-name=dist_julia
#SBATCH --output=%x_%j.out
#SBATCH --error=%x_%j.err

module purge
module load lang/Julia


julia --procs $SLURM_CPUS_PER_TASK parallel_julia_example.jl
</file>

<code bash>
cat dist_julia*.out
</code>

<code bash>
[ ... ]
With 2 CPUs per Task the calculation took 34.029345377 seconds.
</code>

^ CPUs per Task ^ Runtime (s) ^
| 2             | 34.03    |
| 4             | 19.27    |
| 6             | 14.58    |
| 8             | 12.02    |

===== Using MPI with Julia =====
 Of course, you can also use MPI with Julia on MOGON. This requires you to first carry out the following setup for Julia and the MPI interface ''MPI.jl''. 


=== Julia MPI Setup ===

First, Julia must be configured for the use of MPI. For this purpose the [[https://juliaparallel.github.io/MPI.jl/stable/|MPI Wrapper]] for Julia is used.
Log in to one ouf our Service-Nodes and then load Julia and the desired MPI module via:

<code bash>
module load mpi/OpenMPI/4.0.3-GCC-9.3.0
module load lang/Julia
</code>
Next, you need to build the MPI package for Julia with ''Pkg'':

<code bash>
julia -e 'ENV["JULIA_MPI_BINARY"]="system"; using Pkg; Pkg.add("MPI"); Pkg.build("MPI", verbose=true)'
</code>

The output should should be similar to the following if installation and build was successful:

<code julia>
[ ... ]
[ Info: using system MPI                             ]  0/1
┌ Info: Using implementation
│   libmpi = "libmpi"
│   mpiexec_cmd = `mpiexec`
└   MPI_LIBRARY_VERSION_STRING = "Open MPI v4.0.3, package: Open MPI henkela@z0590.mogon Distribution, ident: 4.0.3, repo rev: v4.0.3, Mar 03, 2020\0"
┌ Info: MPI implementation detected
│   impl = OpenMPI::MPIImpl = 2
│   version = v"4.0.3"
└   abi = "OpenMPI"

1 dependency successfully precompiled in 3 seconds (140 already precompiled, 1 skipped during auto due to previous errors)
</code>

=== Julia MPI Example ===
Now that MPI and Julia have been set up correctly, we can proceed to the example.

<file julia hello_mogon_mpi.jl>
using MPI

MPI.Init()
comm = MPI.COMM_WORLD
my_rank = MPI.Comm_rank(comm)
comm_size = MPI.Comm_size(comm)

println("Hello MOGON! I am Rank ", my_rank, " of ", comm_size, " on ", gethostname())

MPI.Finalize()
</file>

<file bash julia_mpi_job.slurm>
#!/bin/bash
#SBATCH --partition=parallel
#SBATCH --account=<YourHPC-Account>
#SBATCH --time=0-00:02:00
#SBATCH --mem-per-cpu=2048 #2GB
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=8
#SBATCH --job-name=mpi_julia
#SBATCH --output=%x_%j.out
#SBATCH --error=%x_%j.err

module purge
module load mpi/OpenMPI/4.0.5-GCC-10.2.0
module load lang/Julia

export JULIA_MPI_PATH=$EBROOTOPENMPI

srun julia -- hello_mogon_mpi.jl
</file>

<code bash>
cat mpi_julia_*.out
</code>

<code bash>
Hello MOGON! I am Rank 9 of 16 on z0278.mogon
Hello MOGON! I am Rank 2 of 16 on z0277.mogon
Hello MOGON! I am Rank 11 of 16 on z0278.mogon
Hello MOGON! I am Rank 12 of 16 on z0278.mogon
Hello MOGON! I am Rank 15 of 16 on z0278.mogon
Hello MOGON! I am Rank 10 of 16 on z0278.mogon
Hello MOGON! I am Rank 13 of 16 on z0278.mogon
Hello MOGON! I am Rank 8 of 16 on z0278.mogon
Hello MOGON! I am Rank 14 of 16 on z0278.mogon
Hello MOGON! I am Rank 1 of 16 on z0277.mogon
Hello MOGON! I am Rank 3 of 16 on z0277.mogon
Hello MOGON! I am Rank 5 of 16 on z0277.mogon
Hello MOGON! I am Rank 4 of 16 on z0277.mogon
Hello MOGON! I am Rank 0 of 16 on z0277.mogon
Hello MOGON! I am Rank 6 of 16 on z0277.mogon
Hello MOGON! I am Rank 7 of 16 on z0277.mogon
</code>


====== Submitting a Julia GPU Job ======

Before you start parallelising with Julia on MOGON GPUs, you need to prepare your Julia environemnt for the usage pf GPUs, as we explaind earlier in the Article about [[https://mogonwiki.zdv.uni-mainz.de/dokuwiki/start:development:scripting_languages:julia#pkg_-_cuda|CUDA.jl]]. After successfully setting up ''CUDA.jl'', you can directly start utilising the advantages of GPUs. We have given some examples below to make it easier for you to start using Julia on MOGON GPUs and to somewhat reflect the advantages of GPUs.


=== CPU/GPU Bandwidth and Read+Write Speed  ===

The test estimates how fast data can be sent to and read from the GPU. Since the GPU is plugged into a PCI bus, this largely depends on the speed of the PCI bus as well as many other factors. However, there is also some overhead included in the measurements, in particular the overhead for function calls and array allocation time. However, since these are present in any "real" use of the GPU, it is reasonable to include them. Memory is allocated and data is sent to the GPU using Julia's ''CUDA.jl'' package. Memory is allocated and data is transferred back to CPU memory using Julia's native ''Array()'' function.

The theoretical bandwidth per lane for **PCIe 3.0** is $0.985 GB/s$. For the GTX 1080Ti (**PCIe3 x16**) used in our MOGON GPU nodes the 16-lane slot could theoretical give $15.754 GB/s$.(( This example was taken from the [[https://de.mathworks.com/help/parallel-computing/measuring-gpu-performance.html|MATLAB Help Center]] and adapted.))

<file julia gpu_rw_perf.jl>
using LinearAlgebra
using Plots
pyplot()
using BenchmarkTools
using Printf
using CUDA
using Random

sizes = 2 .^ (14:30);
timeSend = Array{Float64}(undef, length(sizes));
timeGather = Array{Float64}(undef, length(sizes));
sendBandwidth = Array{Float64}(undef, length(sizes));
gatherBandwidth = Array{Float64}(undef, length(sizes));
memoryTimesCPU = Array{Float64}(undef, length(sizes));
memoryTimesGPU = Array{Float64}(undef, length(sizes));
memoryBandwidthGPU = Array{Float64}(undef, length(sizes));
memoryBandwidthCPU = Array{Float64}(undef, length(sizes));

for i = 1:length(sizes)
    GC.gc(true)
    numElements = convert(Int64, sizes[i] / 8);
    cpuData = rand(0:9, (numElements, 1));
    gpuData = CuArray{Float64}(rand(0:9, (numElements, 1)));
    # Time to GPU
    timeSend[i] = CUDA.@elapsed CuArray(cpuData);
    # Time from GPU
    timeGather[i] = CUDA.@elapsed Array(gpuData);
    sendBandwidth[i] = (sizes[i] / timeSend[i] / 1e9);
    gatherBandwidth[i] = (sizes[i] / timeGather[i] / 1e9);
    memoryTimesGPU[i] = CUDA.@elapsed CUDA.@sync gpuData .+ 1;
    memoryBandwidthGPU[i] = 2*(sizes[i] / memoryTimesGPU[i] / 1e9);
    memoryTimesCPU[i] = @elapsed cpuData .+ 1;
    memoryBandwidthCPU[i] = 2*(sizes[i] / memoryTimesCPU[i] / 1e9);
end

@printf("Achieved peak send speed of %.1f GB/s \n", maximum(sendBandwidth))
@printf("Achieved peak gather speed of %.1f GB/s\n", maximum(gatherBandwidth))
@printf("Achieved peak read+write speed on the GPU: %.1f GB/s\n",maximum(memoryBandwidthGPU))
@printf("Achieved peak read+write speed on the CPU: %.1f GB/s\n",maximum(memoryBandwidthCPU))

p1 = plot(
    sizes,
    sendBandwidth,
    lw = 2,
    legend = :topleft,
    xaxis = ("Array Size (bytes)", :log10),
    xlims = (10^4, 10^9),
    frame = true,
    label = string("Send to GPU (Peak:",round(maximum(sendBandwidth),digits=2)," GB/s)");
);
plot!(p1, sizes, gatherBandwidth, lw = 2, label = string("Gather from GPU (Peak:",round(maximum(gatherBandwidth),digits=2)," GB/s)"));
plot!(p1,yaxis = ("Transfer speed (GB/s)"));
plot!(p1,title = ("Data Transfer Bandwidth"));
plot!(p1,minorxgrid = true, ylims = :round);
scatter!(
    [sizes[argmax(sendBandwidth)], sizes[argmax(gatherBandwidth)]],
    [maximum(sendBandwidth), maximum(gatherBandwidth)],
    label = "",
    marker = (10, 0.3, [:blue, :red]),
);

p2 = plot(
    sizes,
    memoryBandwidthGPU,
    lw = 2,
    legend = :topleft,
    xaxis = ("Array Size (bytes)", :log10),
    xlims = (10^4, 10^9),
    frame = true,
    label = string("GPU (Peak:",round(maximum(memoryBandwidthGPU),digits=2)," GB/s)");
);
plot!(p2,sizes, memoryBandwidthCPU, lw = 2, label = string("CPU (Peak:",round(maximum(memoryBandwidthCPU),digits=2)," GB/s)"));
plot!(p2,yaxis = ("Speed (GB/s)"));
plot!(p2,title = ("read+write Bandwidth"));
plot!(p2,minorxgrid = true, ylims = :round);
scatter!(
    [sizes[argmax(memoryBandwidthGPU)], sizes[argmax(memoryBandwidthCPU)]],
    [maximum(memoryBandwidthGPU), maximum(memoryBandwidthCPU)],
    label = "",
    marker = (10, 0.3, [:blue, :red]),
);

p3 = plot(p1,p2, layout = grid(2, 1, widths=[1]), size=(600,700));

savefig(p3, "gpu_rw_perf_result.png")
</file>

The job script is pretty ordinary. In this example, we use only one GPU and start Julia with four threads. To do this, we request one process with four cpus for multithreading:

<file bash julia_gpu_rw_job.slurm>
#!/bin/bash
#SBATCH --account=<YourHPC-Account>
#SBATCH --job-name=gpu_rw
#SBATCH --output=%x_%j.out
#SBATCH --error=%x_%j.err
#SBATCH --partition=m2_gpu
#SBATCH --gres=gpu:1
#SBATCH --time=0-00:10:00
#SBATCH --mem=11550
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=4

module purge
module load lang/SciPy-bundle/2020.03-foss-2020a-Python-3.8.2
module load vis/matplotlib/3.2.1-foss-2020a-Python-3.8.2
module load system/CUDA/11.0.3-GCC-9.3.0
module load lang/Julia

julia --threads 4 gpu_rw.jl
</file>

The job is submitted with the following command

<code bash>
sbatch julia_gpu_rw_job.slurm
</code>
The job will be finished after a few minutes, you can view the output as follows:
<code bash>
cat gpu_rw*.out
</code>
The output should be similar to the following lines:
<code bash>
Achieved peak send speed of 8.9 GB/s
Achieved peak gather speed of 4.4 GB/s
Achieved peak read+write speed on the GPU: 334.0 GB/s
Achieved peak read+write speed on the CPU: 5.7 GB/s
</code>
 The Julia script also generates a plot, which we would like to show here:

{{ :start:development:scripting_languages:gpu_rw_result.png?direct&600 |}}

=== Memory Intensive Operations ===

You might be familiar with this example if you sumbled upon our [[https://mogonwiki.zdv.uni-mainz.de/dokuwiki/start:development:scripting_languages:matlab#submitting_a_matlab_job_on_gpus|MATLAB article]] or read it on purpose. At this point we would simply like to restate what we originally took from the [[https://de.mathworks.com/help/parallel-computing/measuring-gpu-performance.html|MATLAB Help Center]]:

<blockquote>
For operations where the number of floating-point computations performed per element read from or written to memory is high, the memory speed is much less important. In this case the number and speed of the floating-point units is the limiting factor. These operations are said to have high "computational density".

A good test of computational performance is a matrix-matrix multiply. For multiplying two $N \times N$ matrices, the total number of floating-point calculations is
$$ FLOPS(N) = 2N^3 - N^2 $$

Two input matrices are read and one resulting matrix is written, for a total of $3N^2$ elements read or written. This gives a computational density of $(2N - 1)/3$ FLOP/element. Contrast this with plus as used above, which has a computational density of $1/2$ FLOP/element.
<cite>MATLAB Help Center, [[https://de.mathworks.com/help/parallel-computing/measuring-gpu-performance.html#d123e33031|Measuring GPU Performance]]</cite>
</blockquote>

The difference to our MATLAB article is of course the adaptation to native Julia code. But even so, we have made a few alterations due to the use of the Julia language. When defining vectors or arrays, we have purposely chosen ''Float32'', since GPUs are faster when working with ''Float32'' than with ''Float64''. In addition ''CuArrys'' are by default ''Float32'', as well as functions like ''CUDA.rand'' or ''CUDA.zeros''.

<file julia gpu_perf.jl>
using LinearAlgebra
using Plots
pyplot()
using BenchmarkTools
using Printf
using CUDA

sizes = 2 .^ (12:2:28);
N = convert(Array{Int128,1}, sqrt.(sizes))

timeCPU = Vector{Float32}(undef, length(sizes));
timeGPU = Vector{Float32}(undef, length(sizes));
for i = 1:length(sizes)
    # First on the CPU
    An = rand(Float32, N[i], N[i])
    Bn = rand(Float32, N[i], N[i])
    timeCPU[i] = @elapsed An * Bn
    GC.gc(true)
    # Now on the GPU
    Ac = CUDA.rand(N[i], N[i])
    Bc = CUDA.rand(N[i], N[i])
    timeGPU[i] = CUDA.@elapsed Ac * Bc
    GC.gc(true)
    CUDA.reclaim()
end


gflopsCPU = (2 * N .^ 3 - N .^ 2) ./ timeCPU / 1e9;
gflopsGPU = (2 * N .^ 3 - N .^ 2) ./ timeGPU / 1e9;
@printf(
    "Achieved peak calculation rates of %.1f GFLOPS on CPU, %.1f GFLOPS on GPU ",
    maximum(gflopsCPU),
    maximum(gflopsGPU)
)

plot(
    sizes,
    gflopsCPU,
    lw = 2,
    legend = :topleft,
    xaxis = ("Matrix size (numel)", :log10),
    xlims = (10^3, 10^9),
    frame = true,
    label = string(
        "CPU (Max: ",
        round(maximum(gflopsCPU), digits = 2),
        " GFLOPs @Xeon E5-2650v4)",
    ),
);
plot!(
    sizes,
    gflopsGPU,
    lw = 2,
    label = string(
        "GPU (Max: ",
        round(maximum(gflopsGPU), digits = 2),
        " GFLOPs @GTX 1080Ti)",
    ),
);
plot!(yaxis = ("Calculation Rate (GFLOPS)"));
plot!(title = ("Double precision matrix-matrix multiply"));
plot!(minorxgrid = true, ylims = :round);
scatter!(
    [sizes[argmax(gflopsCPU)], sizes[argmax(gflopsGPU)]],
    [maximum(gflopsCPU), maximum(gflopsGPU)],
    label = "",
    marker = (10, 0.3, [:blue, :red]),
);
savefig("gpu_perf_result.png")
</file>

The job script is quite ordinary. In this example, we only use one GPU and start Julia with four threads. For this we request one process with four cpus for multithreading.

<file bash julia_gpu_perf_job.slurm>
#!/bin/bash
#SBATCH --account=<YourHPC-Account>
#SBATCH --job-name=gpu_perf
#SBATCH --output=%x_%j.out
#SBATCH --error=%x_%j.err
#SBATCH --partition=m2_gpu
#SBATCH --gres=gpu:1
#SBATCH --time=0-00:10:00
#SBATCH --mem=8192
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=4

module purge
module load lang/SciPy-bundle/2020.03-foss-2020a-Python-3.8.2
module load vis/matplotlib/3.2.1-foss-2020a-Python-3.8.2
module load system/CUDA/11.0.3-GCC-9.3.0
module load lang/Julia

julia --threads 4 gpu_perf.jl
</file>

You can submit the job by simply executing:

<code bash>
sbatch julia_gpu_perf_job.slurm
</code>

The job will be completed after acouple of minutes and you can view the output with:


<code bash>
cat gpu_perf*.out
</code>

The Output should resemble the following lines:

<code bash>
Achieved peak calculation rates of 140.4 GFLOPS on CPU, 10134.1 GFLOPS on GPU
</code>

The graphic generated in the script is shown below:

{{ :start:development:scripting_languages:gpu_perf_result_julia.png?direct&600 |}}