archer-gpu-course

Device Management

Available devices

The number of GPUs currently available to the host can be obtained via the API call

  int ndevice = -1;
  cudaGetDeviceCount(&ndevice);

Note that the total number visible may be controlled ultimately by the scheduling system, or some other external consideration.

Devices are numbered logically 0,1,2, ..., ndevice-1. The identity of the currently ‘set’ device, or current context, is

  int myid = -1;
  cudaGetDevice(&myid);

This will be 0 if there is one device. All CUDA API calls, and kernel launches then involve this device.

More then one device

If we have more than one GPU available to the host process, then cudaGetDeviceCount() will return the appropriate number. The initial context will still be device 0, the default.

We can make use of the other devices by switching context with, e.g.,

  int myid1 = 1;
  cudaSetDevice(myid1);

An API call will then refer to the new device. E.g.,

  double * d_data1 = NULL;
  cudaMalloc(&d_data1, ndata*sizeof(double));

will allocate memory on the current device.

Managed memory is slightly different. The CUDA runtime will keep track of what is required where.

The same is true for kernels: a kernel is launched on the current device.

Management

Peer Access

If one has two memory allocations on the same GPU it is perfectly valid to do:

  cudaMemcpy(d_ptr1, d_ptr2, sz, cudaMemcpyDeviceToDevice);

which is a copy within device memory.

More recent NVIDIA (and AMD) devices provide additional fast links between GPUs within a node. (NVIDIA refer to this as NVLINK.) These bypass the need to transfer data via the host (or run a kernel).

PICTURE

This is referred to as “peer access” in the CUDA API.

Querying capability

In general, one should ensure peer access via:

  cudaDeviceCanAccessPeer(int * canAccessPeer, int device1, int device2);

where device1 is the destination device, and device2 is the source device.

If available, it is possible to disable and enable the peer access using

  cudaDeviceDisablePeerAccess(int peerDevice);
  cudaDeviceEnablePeerAccess(int perrDevice, unsigned int flags);

(flags is always set to zero).

When possible, an enabled the relevant link should be used for cudaMemcpyDeviceToDevice copies.

Exercise

Write a simple program which allocates a large array (at least 10 MB, say) on each of two devices using cudaMalloc() (the same size on each device). By making repeated copies of the array with cudaMemcpy(), try to assess the bandwidth which can be obtained by

1. coping from host to device, and then from device to host;
2. copying directly from one device to another using cudaMemcpyDeviceToDevice with peer access disabled;
3. repeating with peer access enabled.

Note that we will need to adjust our queue submission script to ensure that two GPUs are available to the program.

This site is open source. Improve this page.