JM

Table of Contents

Case Study: Monte-Carlo Simulations

Figure 1: Shows an example of a continuous space, discrete time random walk, starting at the origin.

Here, we will provide an example of porting a Monte-Carlo random walk simulation to the GPU, such as the one shown in Figure 1. As with the earlier advice in this course, we will write a function that performs a single step in the random walk.

julia
function mc_random_walk_step(x, sigma)
    return x + randn(typeof(x)) * sigma
end

Now imagine that we want to study some population level statistics, by running many of these walkers in parallel. Despite us choosing a very simple example, Monte-Carlo simulations are an extremely useful tool for computational science. Our aim for this exercise is to perform some number of steps of this Monte-Carlo update for many independent walkers and obtain an array with their final positions in.

We can implement a non-allocating array version of our desired algorithm:

julia
function mc_random_walk!(y, x, sigma, steps)
    # Copy the initial values from x into y
    y .= x
    for t in 1:steps
        y .= mc_random_walk_step.(y, sigma)
    end
    return nothing
end

We can run our algorithm on the CPU easily:

julia
n=2048; x = zeros(Float32, n); y = similar(x);
sigma = 1.0f0; steps=100;
mc_random_walk!(y, x, sigma, steps);

We can extend this to run on the GPU just by changing the types:

julia
x_gpu = to_gpu(x); y_gpu = similar(x_gpu);
mc_random_walk!(y_gpu, x_gpu, sigma, steps);

Let’s benchmark the CPU version:

julia
display(@benchmark mc_random_walk!($y, $x, $sigma, $steps))
Output
BenchmarkTools.Trial: 4828 samples with 1 evaluation per sample.
 Range (min … max):  898.110 μs …   2.498 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     963.186 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):     1.035 ms ± 145.548 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

    ▃▇██▆▅▆▆▅▃▂▁▁▁  ▁       ▁               ▁▄▅▅▅▃▂▁            ▂
  ▆██████████████████████▇▇▇██▇▇█▇█▇▇▇▇▇▇▇▇▇███████████▆▅█▇▄▆▃▄ █
  898 μs        Histogram: log(frequency) by time       1.39 ms <

 Memory estimate: 0 bytes, allocs estimate: 0.

And the GPU version:

julia
display(@benchmark begin
    mc_random_walk!($y_gpu, $x_gpu, $sigma, $steps)
    gpu_synchronize($y_gpu)
end)
Output
BenchmarkTools.Trial: 3143 samples with 1 evaluation per sample.
 Range (min … max):  1.336 ms …   7.695 ms  ┊ GC (min … max): 0.00% … 65.61%
 Time  (median):     1.496 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   1.590 ms ± 417.815 μs  ┊ GC (mean ± σ):  1.42% ±  4.72%

    ▁ ▁█▆       ▁▃                                             
  ▃▇█████▆▅▄▄▄▄▇███▅▄▃▃▃▃▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▁▂▂▂▂▂▂▂▁▂▁▂▂▂▂▁▂▂▂ ▃
  1.34 ms         Histogram: frequency by time        2.62 ms <

 Memory estimate: 295.11 KiB, allocs estimate: 8081.

We see that our GPU version is actually faster (for this number of walkers), however, there is actually a bit of a performance bug under hood that should be addressed. Upon each of our for loops, we are calling a new kernel. Instead, we should try to fuse these kernels together:

julia
function mc_random_walk(x, sigma, steps)
    for t in 1:steps
        x = mc_random_walk_step(x, sigma)
    end
    return x
end

function mc_random_walk_fused!(y, x, sigma, steps)
    y .= mc_random_walk.(x, sigma, steps)
    return nothing
end

Now we can try and benchmark this again:

julia
display(@benchmark begin
    mc_random_walk_fused!($y_gpu, $x_gpu, $sigma, $steps)
    gpu_synchronize($y_gpu)
end)
Output
BenchmarkTools.Trial: 10000 samples with 1 evaluation per sample.
 Range (min … max):  230.100 μs …  3.355 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     377.750 μs              ┊ GC (median):    0.00%
 Time  (mean ± σ):   353.879 μs ± 66.223 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

         ▁                          ▄▇█▅▃▁                      
  ▁▁▂▂▂▃▆██▆▄▄▅▅▄▄▄▄▃▃▃▃▃▃▂▂▂▃▃▄▄▄▅████████▆▅▄▂▂▁▂▁▁▁▁▁▁▁▁▁▁▁▁ ▃
  230 μs          Histogram: frequency by time          490 μs <

 Memory estimate: 3.02 KiB, allocs estimate: 81.

We have fused all the kernel calls together. This minimises the amount of overhead for scheduling, and allows the GPU cores to stay busy for the duration of the computation. This small change allowed us to dramatically improve the performance of our GPU code.

If the kernel calls are very large (e.g. a large matrix multiply), the overhead in calling multiple kernels is only a very small portion of the total time.

Custom Kernel

It is a good exercise to write a custom kernel and compare the execution to the broadcasted notation. Our kernel will be straightforward:

julia
@kernel function mc_random_walk_kernel!(y, x, increments, sigma, steps)
    i = @index(Global, Linear)
    if i <= length(y)
        @inbounds pos = x[i]
        for t in 1:steps
            @inbounds pos += increments[i, t] * sigma
        end
        @inbounds y[i] = pos
    end
end

function mc_random_walk_gpu!(y, x, sigma, steps)
    @assert length(y) == length(x)
    increments = to_gpu(randn(eltype(x), length(x), steps))
    backend = get_backend(y)
    mc_random_walk_kernel!(backend, 256)(y, x, increments, sigma, steps; ndrange=length(y))
    return nothing
end

We can now benchmark on the same data:

julia
display(@benchmark begin
    mc_random_walk_gpu!($y_gpu, $x_gpu, $sigma, $steps)
    gpu_synchronize($y_gpu)
end)
Output
BenchmarkTools.Trial: 4253 samples with 1 evaluation per sample.
 Range (min … max):  948.450 μs …   3.572 ms  ┊ GC (min … max): 0.00% … 31.77%
 Time  (median):       1.113 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):     1.175 ms ± 396.332 μs  ┊ GC (mean ± σ):  4.09% ±  7.69%

  █▇▆▅█▇▆▅▄▂▁▁                                                  ▂
  ████████████▇▇▆▆▅▆▃▅▃▁▃▃▁▁▄▃▄▄▁▃▁▁▁▁▁▃▁▁▁▃▁▁▄▁▁▅▁▅▆▇▅▇▇▇▇█▇██ █
  948 μs        Histogram: log(frequency) by time       3.09 ms <

 Memory estimate: 802.62 KiB, allocs estimate: 74.

We can see that our custom kernel did not perform as well as our simpler approach. It is clear that writing a custom kernel is not necessary to achieve performance gains, as we can rely on the compiler to generate fast code for the GPU, using the array notation. It is entirely possible to rewrite our kernel to be of a similar performance to our previous implementation, but this would require more effort on our part.