Data-Parallel Multithreaded, Distributed, and Multi-GPU Batching

DiffEqFlux.jl allows for data-parallel batching optimally on one computer, across an entire compute cluster, and batching along GPUs. This can be done by parallelizing within an ODE solve or between the ODE solves. The automatic differentiation tooling is compatible with the parallelism. The following examples demonstrate training over a few different modes of parallelism. These examples are not exhaustive.

Within-ODE Multithreaded and GPU Batching

We end by noting that there is an alternative way of batching which can be more efficient in some cases like neural ODEs. With a neural networks, columns are treated independently (by the properties of matrix multiplication). Thus for example, with FastChain we can define an ODE:

using DiffEqFlux, DifferentialEquations

dudt = FastChain(FastDense(2,50,tanh),FastDense(50,2))
p = initial_params(dudt)
f(u,p,t) = dudt(u,p)

and we can solve this ODE where the initial condition is a vector:

u0 = Float32[2.; 0.]
prob = ODEProblem(f,u0,(0f0,1f0),p)
solve(prob,Tsit5())

or we can solve this ODE where the initial condition is a matrix, where each column is an independent system:

u0 = Float32.([0 1 2
               0 0 0])
prob = ODEProblem(f,u0,(0f0,1f0),p)
solve(prob,Tsit5())

On the CPU this will multithread across the system (due to BLAS) and on GPUs this will parallelize the operations across the GPU. To GPU this, you'd simply move the parameters and the initial condition to the GPU:

xs = Float32.([0 1 2
               0 0 0])
prob = ODEProblem(f,gpu(u0),(0f0,1f0),gpu(p))
solve(prob,Tsit5())

This method of parallelism is optimal if all of the operations are linear algebra operations such as a neural ODE. Thus this method of parallelism is demonstrated in the MNIST tutorial.

However, this method of parallelism has many limitations. First of all, the ODE function is required to be written in a way that is independent across the columns. Not all ODEs are written like this, so one needs to be careful. But additionally, this method is ineffective if the ODE function has many serial operations, like u[1]*u[2] - u[3]. In such a case, this indexing behavior will dominate the runtime and cause the parallelism to sometimes even be detrimental.

Out of ODE Parallelism

Instead of parallelizing within an ODE solve, one can parallelize the solves to the ODE itself. While this will be less effective on very large ODEs, like big neural ODE image classifiers, this method be effective even if the ODE is small or the f function is not well-parallelized. This kind of parallelism is done via the DifferentialEquations.jl ensemble interface. The following examples showcase multithreaded, cluster, and (multi)GPU parallelism through this interface.

Multithreaded Batching At a Glance

The following is a full copy-paste example for the multithreading. Distributed and GPU minibatching are described below.

using DifferentialEquations, DiffEqFlux
pa = [1.0]
u0 = [3.0]
θ = [u0;pa]

function model1(θ,ensemble)
  prob = ODEProblem((u, p, t) -> 1.01u .* p, [θ[1]], (0.0, 1.0), [θ[2]])

  function prob_func(prob, i, repeat)
    remake(prob, u0 = 0.5 .+ i/100 .* prob.u0)
  end

  ensemble_prob = EnsembleProblem(prob, prob_func = prob_func)
  sim = solve(ensemble_prob, Tsit5(), ensemble, saveat = 0.1, trajectories = 100)
end

# loss function
loss_serial(θ)   = sum(abs2,1.0.-Array(model1(θ,EnsembleSerial())))
loss_threaded(θ) = sum(abs2,1.0.-Array(model1(θ,EnsembleThreads())))

cb = function (θ,l) # callback function to observe training
  @show l
  false
end

opt = ADAM(0.1)
l1 = loss_serial(θ)
res_serial = DiffEqFlux.sciml_train(loss_serial, θ, opt; cb = cb, maxiters=100)
res_threads = DiffEqFlux.sciml_train(loss_threaded, θ, opt; cb = cb, maxiters=100)

Multithreaded Batching In-Depth

In order to make use of the ensemble interface, we need to build an EnsembleProblem. The prob_func is the function for determining the different DEProblems to solve. This is the place where we can randomly sample initial conditions or pull initial conditions from an array of batches in order to perform our study. To do this, we first define a prototype DEProblem. Here we use the following ODEProblem as our base:

prob = ODEProblem((u, p, t) -> 1.01u .* p, [θ[1]], (0.0, 1.0), [θ[2]])

In the prob_func we define how to build a new problem based on the base problem. In this case, we want to change u0 by a constant, i.e. 0.5 .+ i/100 .* prob.u0 for different trajectories labelled by i. Thus we use the remake function from the problem interface to do so:

function prob_func(prob, i, repeat)
  remake(prob, u0 = 0.5 .+ i/100 .* prob.u0)
end

We now build the EnsembleProblem with this basis:

ensemble_prob = EnsembleProblem(prob, prob_func = prob_func)

Now to solve an ensemble problem, we need to choose an ensembling algorithm and choose the number of trajectories to solve. Here let's solve this in serial with 100 trajectories. Note that i will thus run from 1:100.

sim = solve(ensemble_prob, Tsit5(), EnsembleSerial(), saveat = 0.1, trajectories = 100)

and thus running in multithreading would be:

sim = solve(ensemble_prob, Tsit5(), EnsembleThreads(), saveat = 0.1, trajectories = 100)

This whole mechanism is differentiable, so we then put it in a training loop and it soars. Note that you need to make sure that Julia's multithreading is enabled, which you can do via:

Threads.nthreads()

Distributed Batching Across a Cluster

Changing to distributed computing is very simple as well. The setup is all the same, except you utilize EnsembleDistributed as the ensembler:

sim = solve(ensemble_prob, Tsit5(), EnsembleDistributed(), saveat = 0.1, trajectories = 100)

Note that for this to work you need to ensure that your processes are already started. For more information on setting up processes and utilizing a compute cluster, see the official distributed documentation. The key feature to recognize is that, due to the message passing required for cluster compute, one needs to ensure that all of the required functions are defined on the worker processes. The following is a full example of a distributed batching setup:

using Distributed
addprocs(4)

@everywhere begin
  using DifferentialEquations, DiffEqFlux
  function f(u,p,t)
    1.01u .* p
  end
end

pa = [1.0]
u0 = [3.0]
θ = [u0;pa]

function model1(θ,ensemble)
  prob = ODEProblem(f, [θ[1]], (0.0, 1.0), [θ[2]])

  function prob_func(prob, i, repeat)
    remake(prob, u0 = 0.5 .+ i/100 .* prob.u0)
  end

  ensemble_prob = EnsembleProblem(prob, prob_func = prob_func)
  sim = solve(ensemble_prob, Tsit5(), ensemble, saveat = 0.1, trajectories = 100)
end

cb = function (θ,l) # callback function to observe training
  @show l
  false
end

opt = ADAM(0.1)
loss_distributed(θ) = sum(abs2,1.0.-Array(model1(θ,EnsembleDistributed())))
l1 = loss_distributed(θ)
res_distributed = DiffEqFlux.sciml_train(loss_distributed, θ, opt; cb = cb, maxiters=100)

And note that only addprocs(4) needs to be changed in order to make this demo run across a cluster. For more information on adding processes to a cluster, check out ClusterManagers.jl.

Minibatching Across GPUs with DiffEqGPU

DiffEqGPU.jl allows for generating code parallelizes an ensemble on generated CUDA kernels. This method is efficient for sufficiently small (<100 ODE) problems where the significant computational cost is due to the large number of batch trajectories that need to be solved. This kernel-building process adds a few restrictions to the function, such as requiring it has no boundschecking or allocations. The following is an example of minibatch ensemble parallelism across a GPU:

using DifferentialEquations, DiffEqFlux
function f(du,u,p,t)
  @inbounds begin
    du[1] = 1.01 * u[1] * p[1] * p[2]
  end
end

pa = [1.0]
u0 = [3.0]
θ = [u0;pa]

function model1(θ,ensemble)
  prob = ODEProblem(f, [θ[1]], (0.0, 1.0), [θ[2]])

  function prob_func(prob, i, repeat)
    remake(prob, u0 = 0.5 .+ i/100 .* prob.u0)
  end

  ensemble_prob = EnsembleProblem(prob, prob_func = prob_func)
  sim = solve(ensemble_prob, Tsit5(), ensemble, saveat = 0.1, trajectories = 100)
end

cb = function (θ,l) # callback function to observe training
  @show l
  false
end

opt = ADAM(0.1)
loss_gpu(θ) = sum(abs2,1.0.-Array(model1(θ,EnsembleGPUArray())))
l1 = loss_gpu(θ)
res_gpu = DiffEqFlux.sciml_train(loss_gpu, θ, opt; cb = cb, maxiters=100)

Multi-GPU Batching

DiffEqGPU supports batching across multiple GPUs. See its README for details on setting it up.