This PR implements a performance optimization for DeviceMemoryResource by configuring the memory pool release threshold, addressing slow allocation performance when there are no active allocations.

Problem

As reported in issue #756, allocation via DeviceMemoryResource can be very slow when there are no active allocations. This occurs because the default memory pool is configured with a release threshold of 0, meaning memory is immediately released back to the OS when there are no active suballocations. Subsequent allocations then require expensive OS memory requests.

Solution

This PR upstreams the workaround from CCCL to cuda-core by modifying DeviceMemoryResource.__init__() to:

1. Check the current release threshold using cuMemPoolGetAttribute
2. If the threshold is 0 (default), set it to maximum (0xFFFFFFFFFFFFFFFF) using cuMemPoolSetAttribute
3. This retains memory in the pool for better allocation performance in steady-state scenarios

Implementation Details

• Core change: 22 lines added to DeviceMemoryResource.__init__() in cuda_core/cuda/core/experimental/_memory.py
• Error handling: Wrapped in try-except for backward compatibility with older CUDA drivers
• Documentation: Added comprehensive comments explaining the optimization
• Testing: Added test case to verify proper initialization
• Example: Created performance demonstration script

The implementation follows the exact same pattern as the CCCL workaround, using identical CUDA driver APIs and constants.

Example Usage

from cuda.core.experimental import Device, DeviceMemoryResource

device = Device()
mr = DeviceMemoryResource(device.device_id)  # Now automatically optimized

# Subsequent allocations will be faster due to memory retention
buffer = mr.allocate(1024)
buffer.close()

The optimization is transparent to users and maintains full backward compatibility.

Fixes #771.

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.