[FEA] The libcudf example should show how to use a memory resource

[jrhemstad]

Is your feature request related to a problem? Please describe.

Using a memory pool is essential to good performance with libcudf. Today, the example libcudf application doesn't show how to construct an RMM memory resource nor set the current_device_resource.

Describe the solution you'd like

The example should be updated to show constructing a rmm::mr::pool_memory_resource or rmm::mr::cuda_async_memory_resource and setting it as the default via set_current_device_resource

Describe alternatives you've considered
One could argue that this could/should be a separate example, but using a pool resource is fundamental to performance in libcudf and makes sense to be part of the existing and future examples.

[davidwendt]

I could add some more comments to the example in #11849
https://github.com/rapidsai/cudf/pull/11849/files#diff-5c4769eb07f9859a4f196ad1fc2105ef1420f1cc645401eb9c4f27c601730cefR70-R74
Do you think that would be enough or do you have something else in mind?

[jrhemstad]

Here's what I'm thinking:

int main(int argc, char** argv)
{

  // Construct a memory pool using RAPIDS Memory Manager (RMM) 
  // Using a memory pool for device memory allocations is important for good performance in libcudf
  // Defaults to allocating half of the available GPU memory 
  rmm::mr::pool_memory_resource mr{};

  // Set the pool resource to be used by default for all device memory allocations
  // Note: It is the user's responsibility to ensure the `mr` object stays alive for the duration of it being set as the default
  rmm::mr::set_current_device_resource(&mr);

  // Read data
  auto stock_table_with_metadata = read_csv("4stock_5day.csv");

  // Process
  auto result = average_closing_price(*stock_table_with_metadata.tbl);

  // Write out result
  write_csv(*result, "4stock_5day_avg_close.csv");

  return 0;
}