Memory Management

CuPy uses memory pool for memory allocations by default. The memory pool significantly improves the performance by mitigating the overhead of memory allocation and CPU/GPU synchronization.

There are two different memory pools in CuPy:

  • Device memory pool (GPU device memory), which is used for GPU memory allocations.

  • Pinned memory pool (non-swappable CPU memory), which is used during CPU-to-GPU data transfer.

Attention

When you monitor the memory usage (e.g., using nvidia-smi for GPU memory or ps for CPU memory), you may notice that memory not being freed even after the array instance become out of scope. This is an expected behavior, as the default memory pool “caches” the allocated memory blocks.

See Low-Level CUDA Support for the details of memory management APIs.

Memory Pool Operations

The memory pool instance provides statistics about memory allocation. To access the default memory pool instance, use cupy.get_default_memory_pool() and cupy.get_default_pinned_memory_pool(). You can also free all unused memory blocks hold in the memory pool. See the example code below for details:

import cupy
import numpy

mempool = cupy.get_default_memory_pool()
pinned_mempool = cupy.get_default_pinned_memory_pool()

# Create an array on CPU.
# NumPy allocates 400 bytes in CPU (not managed by CuPy memory pool).
a_cpu = numpy.ndarray(100, dtype=numpy.float32)
print(a_cpu.nbytes)                      # 400

# You can access statistics of these memory pools.
print(mempool.used_bytes())              # 0
print(mempool.total_bytes())             # 0
print(pinned_mempool.n_free_blocks())    # 0

# Transfer the array from CPU to GPU.
# This allocates 400 bytes from the device memory pool, and another 400
# bytes from the pinned memory pool.  The allocated pinned memory will be
# released just after the transfer is complete.  Note that the actual
# allocation size may be rounded to larger value than the requested size
# for performance.
a = cupy.array(a_cpu)
print(a.nbytes)                          # 400
print(mempool.used_bytes())              # 512
print(mempool.total_bytes())             # 512
print(pinned_mempool.n_free_blocks())    # 1

# When the array goes out of scope, the allocated device memory is released
# and kept in the pool for future reuse.
a = None  # (or `del a`)
print(mempool.used_bytes())              # 0
print(mempool.total_bytes())             # 512
print(pinned_mempool.n_free_blocks())    # 1

# You can clear the memory pool by calling `free_all_blocks`.
mempool.free_all_blocks()
pinned_mempool.free_all_blocks()
print(mempool.used_bytes())              # 0
print(mempool.total_bytes())             # 0
print(pinned_mempool.n_free_blocks())    # 0

See cupy.cuda.MemoryPool and cupy.cuda.PinnedMemoryPool for details.

Limiting GPU Memory Usage

You can hard-limit the amount of GPU memory that can be allocated by using CUPY_GPU_MEMORY_LIMIT environment variable (see Environment variables for details).

# Set the hard-limit to 1 GiB:
#   $ export CUPY_GPU_MEMORY_LIMIT="1073741824"

# You can also specify the limit in fraction of the total amount of memory
# on the GPU. If you have a GPU with 2 GiB memory, the following is
# equivalent to the above configuration.
#   $ export CUPY_GPU_MEMORY_LIMIT="50%"

import cupy
print(cupy.get_default_memory_pool().get_limit())  # 1073741824

You can also set the limit (or override the value specified via the environment variable) using cupy.cuda.MemoryPool.set_limit(). In this way, you can use a different limit for each GPU device.

import cupy

mempool = cupy.get_default_memory_pool()

with cupy.cuda.Device(0):
    mempool.set_limit(size=1024**3)  # 1 GiB

with cupy.cuda.Device(1):
    mempool.set_limit(size=2*1024**3)  # 2 GiB

Note

CUDA allocates some GPU memory outside of the memory pool (such as CUDA context, library handles, etc.). Depending on the usage, such memory may take one to few hundred MiB. That will not be counted in the limit.

Changing Memory Pool

You can use your own memory allocator instead of the default memory pool by passing the memory allocation function to cupy.cuda.set_allocator() / cupy.cuda.set_pinned_memory_allocator(). The memory allocator function should take 1 argument (the requested size in bytes) and return cupy.cuda.MemoryPointer / cupy.cuda.PinnedMemoryPointer.

CuPy provides two such allocators for using managed memory and stream ordered memory on GPU, see cupy.cuda.malloc_managed() and cupy.cuda.malloc_async(), respectively, for details. To enable a memory pool backed by managed memory, you can construct a new MemoryPool instance with its allocator set to malloc_managed() as follows

import cupy

# Use managed memory
cupy.cuda.set_allocator(cupy.cuda.MemoryPool(cupy.cuda.malloc_managed).malloc)

Note that if you pass malloc_managed() directly to set_allocator() without constructing a MemoryPool instance, when the memory is freed it will be released back to the system immediately, which may or may not be desired.

Stream Ordered Memory Allocator is a new feature added since CUDA 11.2. CuPy provides an experimental interface to it. Similar to CuPy’s memory pool, Stream Ordered Memory Allocator also allocates and deallocates memory asynchronously from a memory pool in a stream-ordered fashion. The key difference is that it is a built-in feature provided in CUDA by NVIDIA. To enable a memory pool that manages stream ordered memory, you can do

import cupy

# Use asynchronous stream ordered memory
cupy.cuda.set_allocator(cupy.cuda.malloc_async)

# Create a custom stream
s = cupy.cuda.Stream()

# This would allocate memory asynchronously on stream s
with s:
    a = cupy.empty((100,), dtype=cupy.float64)

Note that in this case we do not create a new MemoryPool instance, as the pool is directly managed by the CUDA driver. (Currently CuPy uses the device’s default memory pool, but it is not guaranteed.)

When using stream ordered memory, it is important that you maintain a correct stream semantics yourselves using, for example, the Stream and Event APIs (see Low-Level CUDA Support for details); CuPy does not attempt to act smartly for you. Upon deallocation, the memory is freed asynchronously either on the stream it was allocated (first attempt), or on any current CuPy stream (second attempt). It is permitted that the stream on which the memory was allocated gets destroyed before all memory allocated on it is freed.

You can even disable the default memory pool by the code below. Be sure to do this before any other CuPy operations.

import cupy

# Disable memory pool for device memory (GPU)
cupy.cuda.set_allocator(None)

# Disable memory pool for pinned memory (CPU).
cupy.cuda.set_pinned_memory_allocator(None)