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.
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.
For using pinned memory more conveniently, we also provide a few high-level APIs in the
cupyx.zeros_like_pinned(). They return NumPy arrays backed by pinned memory. If CuPy’s pinned memory pool
is in use, the pinned memory is allocated from the pool.
CuPy v8 and above provides a FFT plan cache that could use a portion of device memory if FFT and related functions are used. The memory taken can be released by shrinking or disabling the cache.
Memory Pool Operations#
The memory pool instance provides statistics about memory allocation.
To access the default memory pool instance, use
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
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
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
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
The memory allocator function should take 1 argument (the requested size in bytes) and return
CuPy provides two such allocators for using managed memory and stream ordered memory on GPU,
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
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
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/deallocates memory asynchronously from/to a memory pool in a stream-ordered fashion. The key difference is that it is a built-in feature implemented in the CUDA driver by NVIDIA, so other CUDA applications in the same processs can easily allocate memory from the same pool.
To enable a memory pool that manages stream ordered memory, you can construct a new
import cupy # Use asynchronous stream ordered memory cupy.cuda.set_allocator(cupy.cuda.MemoryAsyncPool().malloc) # 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 use the
MemoryPool class. The
a different input argument from that of
MemoryPool to indicate which pool to use.
Please refer to
MemoryAsyncPool’s documentation for further detail.
Note that if you pass
malloc_async() directly to
set_allocator() without constructing
MemoryAsyncPool instance, the device’s current memory pool will be used.
When using stream ordered memory, it is important that you maintain a correct stream semantics yourselves using, for example,
Event APIs (see Streams and Events 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.
In addition, applications/libraries internally use
cudaMalloc (CUDA’s default, synchronous allocator) could have unexpected
interplay with Stream Ordered Memory Allocator. Specifically, memory freed to the memory pool might not be immediately visible
cudaMalloc, leading to potential out-of-memory errors. In this case, you can either call
or just manually perform a (event/stream/device) synchronization, and retry.
MemoryAsyncPool interface is experimental. In particular, while its API is largely identical
to that of
MemoryPool, several of the pool’s methods require a sufficiently new driver (and of course, a
supported hardware, CUDA version, and platform) due to CUDA’s limitation.
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)