# Run MATLAB Functions on Multiple GPUs

This example shows how to run MATLAB® code on multiple GPUs in parallel, first on your local machine, then scaling up to a cluster. As a sample problem, the example uses the logistic map, an equation that models the growth of a population.

A growing number of features in MATLAB offer automatic parallel support, including multi-gpu support, without requiring any extra coding. For details, see Run MATLAB Functions with Automatic Parallel Support. For example, the `trainNetwork` function offers multi-gpu support for training of neural networks and inference. For more information, see Scale Up Deep Learning in Parallel, on GPUs, and in the Cloud (Deep Learning Toolbox).

### Use a Single GPU

To run computations on a single GPU, use `gpuArray` objects as inputs to GPU-enabled MATLAB functions. To learn more about GPU-enabled functions, see Run MATLAB Functions on a GPU.

Create GPU arrays defining the growth rate, `r`, and the population, `x`. For more information on creating `gpuArray` objects, see Establish Arrays on a GPU.

```N = 1000; r = gpuArray.linspace(0,4,N); x = rand(1,N,"gpuArray");```

Use a simple algorithm to iterate the logistic map. Because the algorithm uses GPU-enabled operators on `gpuArray` input data, the computations run on the GPU.

```numIterations = 1000; for n=1:numIterations x = r.*x.*(1-x); end```

When the computations are done, plot the growth rate against the population.

`plot(r,x,'.');` If you need more performance, GPU arrays supports several options. For a list, see the `gpuArray` function page. For example, the algorithm in this example only performs element-wise operations on GPU arrays, and so you can use the `arrayfun` function to precompile them for GPU.

### Use Multiple GPUs with `parfor`

You can use `parfor`-loops to distribute `for`-loop iterations among parallel workers. If your computations use GPU-enabled functions, then the computations run on the GPU of the worker. For example, if you use the Monte Carlo method to randomly simulate the evolution of populations, simulations are computed with multiple GPUs in parallel using a `parfor`-loop.

Create a parallel pool with as many workers as GPUs available using `parpool`. To determine the number of GPUs available, use the `gpuDeviceCount` function. By default, MATLAB assigns a different GPU to each worker for best performance. For more information on selecting GPUs in a parallel pool, see Use Multiple GPUs in Parallel Pool.

```numGPUs = gpuDeviceCount("available"); parpool(numGPUs);```
```Starting parallel pool (parpool) using the 'Processes' profile ... Connected to the parallel pool (number of workers: 2). ```

Define the number of simulations, and create an array in the GPU to store the population vector for each simulation.

```numSimulations = 100; X = zeros(numSimulations,N,"gpuArray");```

Use a `parfor` loop to distribute simulations to workers in the pool. The code inside the loop creates a random `gpuArray` for the initial population, and iterates the logistic map on it. Because the code uses GPU-enabled operators on `gpuArray` input data, the computations automatically run on the GPU of the worker.

```parfor i = 1:numSimulations X(i,:) = rand(1,N,"gpuArray"); for n=1:numIterations X(i,:) = r.*X(i,:).*(1-X(i,:)); end end```

When the computations are done, plot the results of all simulations. Each color represents a different simulation.

```figure plot(r,X,'.');``` With default preferences, `parpool` starts a parallel pool of process workers. Running code in parallel on process workers often results in data being copied to each worker which can use a significant amount of GPU memory when working with GPU arrays. In contrast, thread workers can share memory. To reduce memory usage and lower data transfer costs, use a parallel pool of thread workers by calling `parpool("Threads")`. Thread workers support only a subset of functions available for processes workers. For more information, see Choose Between Thread-Based and Process-Based Environments.

If you need greater control over your calculations, you can use more advanced parallel functionality. For example, you can use a `parallel.pool.DataQueue` to send data from the workers during computations. For an example, see Plot During Parameter Sweep with parfor.

If you want to generate a reproducible set of random numbers, you can control the random number generation on the GPU of a worker. For more information, see Random Number Streams on a GPU.

### Use Multiple GPUs Asynchronously with `parfeval`

You can use `parfeval` to run computations asynchronously on parallel pool workers. If your computations use GPU-enabled functions, then the computations run on the GPU of the worker. As an example, you run Monte Carlo simulations on multiple GPUs asynchronously.

To hold the results of computations after the workers complete them, use future objects. Preallocate an array of future objects for the result of each simulation.

`f(numSimulations) = parallel.FevalFuture;`

To run computations with `parfeval`, you must place them inside a function. For example, `myParallelFcn` contains the code of a single simulation.

`type myParallelFcn`
```function x = myParallelFcn(r) N = 1000; x = gpuArray.rand(1,N); numIterations = 1000; for n=1:numIterations x = r.*x.*(1-x); end end ```

Use a `for` loop to loop over simulations, and use `parfeval` to run them asynchronously on a worker in the parallel pool. `myParallelFcn` uses GPU-enabled functions on `gpuArray` input data, so they run on the GPU of the worker. Because `parfeval` performs the computations asynchronously, it does not block MATLAB, and you can continue working while computations happen.

```for i=1:numSimulations f(i) = parfeval(@myParallelFcn,1,r); end```

To collect the results from `parfeval` when they are ready, you can use `fetchOutputs` or `fetchNext` on the future objects. Also, you can use `afterEach` or `afterAll` to invoke functions on the results automatically when they are ready. For example, to plot the result of each simulation immediately after it completes, use `afterEach` on the future objects. Each color represents a different simulation.

```figure hold on afterEach(f,@(x) plot(r,x,'.'),0);``` ### Use Multiple GPUs in a Cluster

If you have access to a cluster with multiple GPUs, then you can scale up your computations. Use the `parpool` function to start a parallel pool on the cluster. When you do so, parallel features, such as `parfor` loops or `parfeval`, run on the cluster workers. If your computations use GPU-enabled functions on `gpuArray` input data, then those functions run on the GPU of the cluster worker. Note that cluster features are supported only in process-based environments. To learn more about running parallel features on a cluster, see Scale Up from Desktop to Cluster.

### Advanced Support for Fast Multi-Node GPU Communication

Some multi-GPU features in MATLAB®, including `trainNetwork`, are optimized for direct communication via fast interconnects for improved performance.

If you have appropriate hardware connections, then data transfer between multiple GPUs uses fast peer-to-peer communication, including NVLink, if available.

If you are using a Linux compute cluster with fast interconnects between machines such as Infiniband, or fast interconnects between GPUs on different machines, such as GPUDirect RDMA, you might be able to take advantage of fast multi-node support in MATLAB. Enable this support on all the workers in your pool by setting the environment variable `PARALLEL_SERVER_FAST_MULTINODE_GPU_COMMUNICATION` to `1`. Set this environment variable in the Cluster Profile Manager.

This feature is part of the NVIDIA NCCL library for GPU communication. To configure it, you must set additional environment variables to define the network interface protocol, especially `NCCL_SOCKET_IFNAME`. For more information, see the NCCL documentation and in particular the section on NCCL Environment Variables.