Add Distributed Examples to Documentation

[rkierulf]

I added two examples I tested yesterday to the "how-to" section of the documentation, one for running a simulation using multiple GPUs, and another using multiple nodes on a SLURM cluster. This isn't included in the example scripts, but I verified the accuracy compared to a normal simulation with these lines of code added at the end:

signal_2 = simulate(obj, seq, sys)
println([sum(abs.(signal.profiles[i].data .- signal_2.profiles[i].data)) / length(signal.profiles[i].data) for i=1:length(signal.profiles)])

The values I saw were all on the scale of 10e-6 / 10e-7, which is near machine precision for Float32s.

[cncastillo]

https://juliahealth.org/KomaMRI.jl/previews/PR468/

[cncastillo]

btw @rkierulf @pvillacorta I think this could be very useful to simulate the "aorta" with flowing spins, as you can distribute the phantom to multiple nodes or GPUs!!

It would be cool if @pvillacorta could test this and put some photos :D As an example a phantom that was too big to fit in one GPU

[pvillacorta]

Nice! Once #442 is finished, I will try to simulate with the aorta and I will make an example for this :)

[codecov]

Codecov Report

Attention: Patch coverage is 87.50000% with 1 line in your changes missing coverage. Please review.

Project coverage is 90.93%. Comparing base (57f8698) to head (d9f10b9).
Report is 1 commits behind head on master.

Files	Patch %	Lines
KomaMRICore/src/simulation/GPUFunctions.jl	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #468      +/-   ##
==========================================
+ Coverage   90.75%   90.93%   +0.17%     
==========================================
  Files          53       53              
  Lines        2932     2934       +2     
==========================================
+ Hits         2661     2668       +7     
+ Misses        271      266       -5     

Flag	Coverage Δ
base	88.20% <ø> (ø)
core	92.61% <87.50%> (+0.95%)	⬆️
files	93.55% <ø> (ø)
komamri	93.98% <ø> (ø)
plots	89.30% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

Files	Coverage Δ
KomaMRICore/src/rawdata/ISMRMRD.jl	98.50% <100.00%> (+1.78%)	⬆️
KomaMRICore/src/simulation/GPUFunctions.jl	60.52% <0.00%> (-1.64%)	⬇️

... and 1 file with indirect coverage changes

[cncastillo]

What about adding some images? (inspired by https://classic.d2l.ai/chapter_computational-performance/parameterserver.html)

[cncastillo]

This is very nice! quite amazing that it works. Could you:

• add a test for +(sig1, sig2)
• rename the sig's to raw's
• add the images from my previous comment
• add the SLURM script as dropdowns with !!! details "SLURM script" ... (example: https://github.com/JuliaHealth/KomaMRI.jl/blob/master/docs/src/index.md?plain=1)

[github-actions]

KomaMRI Benchmarks

Benchmark suite	Current: c309ab6	Previous: 57f8698	Ratio
MRI Lab/Bloch/CPU/2 thread(s)	227228176 ns	226618506 ns	1.00
MRI Lab/Bloch/CPU/4 thread(s)	174890811.5 ns	174536994 ns	1.00
MRI Lab/Bloch/CPU/8 thread(s)	92125490.5 ns	146360095.5 ns	0.63
MRI Lab/Bloch/CPU/1 thread(s)	340803231.5 ns	347644824 ns	0.98
MRI Lab/Bloch/GPU/CUDA	57453774.5 ns	57253633 ns	1.00
MRI Lab/Bloch/GPU/oneAPI	513802813 ns	515042255.5 ns	1.00
MRI Lab/Bloch/GPU/Metal	587853374.5 ns	541353541 ns	1.09
MRI Lab/Bloch/GPU/AMDGPU	37310330 ns	37619574.5 ns	0.99
Slice Selection 3D/Bloch/CPU/2 thread(s)	1175371127 ns	1024148878 ns	1.15
Slice Selection 3D/Bloch/CPU/4 thread(s)	588597986 ns	580936747 ns	1.01
Slice Selection 3D/Bloch/CPU/8 thread(s)	349205204.5 ns	386777586 ns	0.90
Slice Selection 3D/Bloch/CPU/1 thread(s)	1927153559.5 ns	1925568005.5 ns	1.00
Slice Selection 3D/Bloch/GPU/CUDA	101738262 ns	100754922 ns	1.01
Slice Selection 3D/Bloch/GPU/oneAPI	628750808 ns	654922437.5 ns	0.96
Slice Selection 3D/Bloch/GPU/Metal	563667813 ns	564653500 ns	1.00
Slice Selection 3D/Bloch/GPU/AMDGPU	60715236 ns	60779232 ns	1.00

This comment was automatically generated by workflow using github-action-benchmark.

[cncastillo]

This is what the HPC example is actually doing