Memory (management)

Devices

class arrow::Device : public std::enable_shared_from_this<Device>, public arrow::util::EqualityComparable<Device>

EXPERIMENTAL: Abstract interface for hardware devices.

This object represents a device with access to some memory spaces. When handling a Buffer or raw memory address, it allows deciding in which context the raw memory address should be interpreted (e.g. CPU-accessible memory, or embedded memory on some particular GPU).

Subclassed by arrow::CPUDevice, arrow::cuda::CudaDevice

Public Functions

const char *type_name() const = 0

A shorthand for this device’s type.

The returned value is different for each device class, but is the same for all instances of a given class. It can be used as a replacement for RTTI.

std::string ToString() const = 0

A human-readable description of the device.

The returned value should be detailed enough to distinguish between different instances, where necessary.

bool Equals(const Device&) const = 0

Whether this instance points to the same device as another one.

bool is_cpu() const

Whether this device is the main CPU device.

This shorthand method is very useful when deciding whether a memory address is CPU-accessible.

std::shared_ptr<MemoryManager> default_memory_manager() = 0

Return a MemoryManager instance tied to this device.

The returned instance uses default parameters for this device type’s MemoryManager implementation. Some devices also allow constructing MemoryManager instances with non-default parameters.

class arrow::CPUDevice : public arrow::Device

Public Functions

const char *type_name() const override

A shorthand for this device’s type.

The returned value is different for each device class, but is the same for all instances of a given class. It can be used as a replacement for RTTI.

std::string ToString() const override

A human-readable description of the device.

The returned value should be detailed enough to distinguish between different instances, where necessary.

bool Equals(const Device&) const override

Whether this instance points to the same device as another one.

std::shared_ptr<MemoryManager> default_memory_manager() override

Return a MemoryManager instance tied to this device.

The returned instance uses default parameters for this device type’s MemoryManager implementation. Some devices also allow constructing MemoryManager instances with non-default parameters.

Public Static Functions

std::shared_ptr<Device> Instance()

Return the global CPUDevice instance.

std::shared_ptr<MemoryManager> memory_manager(MemoryPool *pool)

Create a MemoryManager.

The returned MemoryManager will use the given MemoryPool for allocations.

std::shared_ptr<MemoryManager> arrow::default_cpu_memory_manager()

Return the default CPU MemoryManager instance.

The returned singleton instance uses the default MemoryPool. This function is a faster spelling of CPUDevice::Instance()->default_memory_manager().

Memory Managers

class arrow::MemoryManager : public std::enable_shared_from_this<MemoryManager>

EXPERIMENTAL: An object that provides memory management primitives.

A MemoryManager is always tied to a particular Device instance. It can also have additional parameters (such as a MemoryPool to allocate CPU memory).

Subclassed by arrow::CPUMemoryManager, arrow::cuda::CudaMemoryManager

Public Functions

const std::shared_ptr<Device> &device() const

The device this MemoryManager is tied to.

bool is_cpu() const

Whether this MemoryManager is tied to the main CPU device.

This shorthand method is very useful when deciding whether a memory address is CPU-accessible.

Result<std::shared_ptr<io::RandomAccessFile>> GetBufferReader(std::shared_ptr<Buffer> buf) = 0

Create a RandomAccessFile to read a particular buffer.

The given buffer must be tied to this MemoryManager.

See also the Buffer::GetReader shorthand.

Result<std::shared_ptr<io::OutputStream>> GetBufferWriter(std::shared_ptr<Buffer> buf) = 0

Create a OutputStream to write to a particular buffer.

The given buffer must be mutable and tied to this MemoryManager. The returned stream object writes into the buffer’s underlying memory (but it won’t resize it).

See also the Buffer::GetWriter shorthand.

Result<std::shared_ptr<Buffer>> AllocateBuffer(int64_t size) = 0

Allocate a (mutable) Buffer.

The buffer will be allocated in the device’s memory.

Public Static Functions

Result<std::shared_ptr<Buffer>> CopyBuffer(const std::shared_ptr<Buffer> &source, const std::shared_ptr<MemoryManager> &to)

Copy a Buffer to a destination MemoryManager.

See also the Buffer::Copy shorthand.

Result<std::shared_ptr<Buffer>> ViewBuffer(const std::shared_ptr<Buffer> &source, const std::shared_ptr<MemoryManager> &to)

Make a no-copy Buffer view in a destination MemoryManager.

See also the Buffer::View shorthand.

class arrow::CPUMemoryManager : public arrow::MemoryManager

Public Functions

Result<std::shared_ptr<io::RandomAccessFile>> GetBufferReader(std::shared_ptr<Buffer> buf) override

Create a RandomAccessFile to read a particular buffer.

The given buffer must be tied to this MemoryManager.

See also the Buffer::GetReader shorthand.

Result<std::shared_ptr<io::OutputStream>> GetBufferWriter(std::shared_ptr<Buffer> buf) override

Create a OutputStream to write to a particular buffer.

The given buffer must be mutable and tied to this MemoryManager. The returned stream object writes into the buffer’s underlying memory (but it won’t resize it).

See also the Buffer::GetWriter shorthand.

Result<std::shared_ptr<Buffer>> AllocateBuffer(int64_t size) override

Allocate a (mutable) Buffer.

The buffer will be allocated in the device’s memory.

MemoryPool *pool() const

Return the MemoryPool associated with this MemoryManager.

Buffers

class arrow::Buffer

Object containing a pointer to a piece of contiguous memory with a particular size.

Buffers have two related notions of length: size and capacity. Size is the number of bytes that might have valid data. Capacity is the number of bytes that were allocated for the buffer in total.

The Buffer base class does not own its memory, but subclasses often do.

The following invariant is always true: Size <= Capacity

Subclassed by arrow::cuda::CudaBuffer, arrow::MutableBuffer, arrow::py::NumPyBuffer, arrow::py::PyBuffer, arrow::py::PyForeignBuffer

Public Functions

Buffer(const uint8_t *data, int64_t size)

Construct from buffer and size without copying memory.

Note

The passed memory must be kept alive through some other means

Parameters
  • [in] data: a memory buffer

  • [in] size: buffer size

Buffer(util::string_view data)

Construct from string_view without copying memory.

Note

The memory viewed by data must not be deallocated in the lifetime of the Buffer; temporary rvalue strings must be stored in an lvalue somewhere

Parameters
  • [in] data: a string_view object

Buffer(const std::shared_ptr<Buffer> &parent, const int64_t offset, const int64_t size)

An offset into data that is owned by another buffer, but we want to be able to retain a valid pointer to it even after other shared_ptr’s to the parent buffer have been destroyed.

This method makes no assertions about alignment or padding of the buffer but in general we expected buffers to be aligned and padded to 64 bytes. In the future we might add utility methods to help determine if a buffer satisfies this contract.

std::string ToHexString()

Construct a new std::string with a hexadecimal representation of the buffer.

Return

std::string

bool Equals(const Buffer &other, int64_t nbytes) const

Return true if both buffers are the same size and contain the same bytes up to the number of compared bytes.

bool Equals(const Buffer &other) const

Return true if both buffers are the same size and contain the same bytes.

Result<std::shared_ptr<Buffer>> CopySlice(const int64_t start, const int64_t nbytes, MemoryPool *pool = default_memory_pool()) const

Copy a section of the buffer into a new Buffer.

void ZeroPadding()

Zero bytes in padding, i.e. bytes between size_ and capacity_.

std::string ToString() const

Copy buffer contents into a new std::string.

Return

std::string

Note

Can throw std::bad_alloc if buffer is large

operator util::string_view() const

View buffer contents as a util::string_view.

Return

util::string_view

operator util::bytes_view() const

View buffer contents as a util::bytes_view.

Return

util::bytes_view

const uint8_t *data() const

Return a pointer to the buffer’s data.

The buffer has to be a CPU buffer (is_cpu() is true). Otherwise, an assertion may be thrown or a null pointer may be returned.

To get the buffer’s data address regardless of its device, call address().

uint8_t *mutable_data()

Return a writable pointer to the buffer’s data.

The buffer has to be a mutable CPU buffer (is_cpu() and is_mutable() are true). Otherwise, an assertion may be thrown or a null pointer may be returned.

To get the buffer’s mutable data address regardless of its device, call mutable_address().

uintptr_t address() const

Return the device address of the buffer’s data.

uintptr_t mutable_address() const

Return a writable device address to the buffer’s data.

The buffer has to be a mutable buffer (is_mutable() is true). Otherwise, an assertion may be thrown or 0 may be returned.

int64_t size() const

Return the buffer’s size in bytes.

int64_t capacity() const

Return the buffer’s capacity (number of allocated bytes)

bool is_cpu() const

Whether the buffer is directly CPU-accessible.

If this function returns true, you can read directly from the buffer’s data() pointer. Otherwise, you’ll have to View() or Copy() it.

bool is_mutable() const

Whether the buffer is mutable.

If this function returns true, you are allowed to modify buffer contents using the pointer returned by mutable_data() or mutable_address().

Public Static Functions

std::shared_ptr<Buffer> FromString(std::string data)

Construct an immutable buffer that takes ownership of the contents of an std::string (without copying it).

Return

a new Buffer instance

Parameters
  • [in] data: a string to own

template<typename T, typename SizeType = int64_t>
std::shared_ptr<Buffer> Wrap(const T *data, SizeType length)

Create buffer referencing typed memory with some length without copying.

Return

a new shared_ptr<Buffer>

Parameters
  • [in] data: the typed memory as C array

  • [in] length: the number of values in the array

template<typename T>
std::shared_ptr<Buffer> Wrap(const std::vector<T> &data)

Create buffer referencing std::vector with some length without copying.

Return

a new shared_ptr<Buffer>

Parameters
  • [in] data: the vector to be referenced. If this vector is changed, the buffer may become invalid

Result<std::shared_ptr<io::RandomAccessFile>> GetReader(std::shared_ptr<Buffer>)

Get a RandomAccessFile for reading a buffer.

The returned file object reads from this buffer’s underlying memory.

Result<std::shared_ptr<io::OutputStream>> GetWriter(std::shared_ptr<Buffer>)

Get a OutputStream for writing to a buffer.

The buffer must be mutable. The returned stream object writes into the buffer’s underlying memory (but it won’t resize it).

Result<std::shared_ptr<Buffer>> Copy(std::shared_ptr<Buffer> source, const std::shared_ptr<MemoryManager> &to)

Copy buffer.

The buffer contents will be copied into a new buffer allocated by the given MemoryManager. This function supports cross-device copies.

Result<std::shared_ptr<Buffer>> View(std::shared_ptr<Buffer> source, const std::shared_ptr<MemoryManager> &to)

View buffer.

Return a Buffer that reflects this buffer, seen potentially from another device, without making an explicit copy of the contents. The underlying mechanism is typically implemented by the kernel or device driver, and may involve lazy caching of parts of the buffer contents on the destination device’s memory.

If a non-copy view is unsupported for the buffer on the given device, nullptr is returned. An error can be returned if some low-level operation fails (such as an out-of-memory condition).

Result<std::shared_ptr<Buffer>> ViewOrCopy(std::shared_ptr<Buffer> source, const std::shared_ptr<MemoryManager> &to)

View or copy buffer.

Try to view buffer contents on the given MemoryManager’s device, but fall back to copying if a no-copy view isn’t supported.

class arrow::MutableBuffer : public arrow::Buffer

A Buffer whose contents can be mutated.

May or may not own its data.

Subclassed by arrow::cuda::CudaHostBuffer, arrow::ResizableBuffer

Public Static Functions

template<typename T, typename SizeType = int64_t>
std::shared_ptr<Buffer> Wrap(T *data, SizeType length)

Create buffer referencing typed memory with some length.

Return

a new shared_ptr<Buffer>

Parameters
  • [in] data: the typed memory as C array

  • [in] length: the number of values in the array

class arrow::ResizableBuffer : public arrow::MutableBuffer

A mutable buffer that can be resized.

Public Functions

Status Resize(const int64_t new_size, bool shrink_to_fit = true) = 0

Change buffer reported size to indicated size, allocating memory if necessary.

This will ensure that the capacity of the buffer is a multiple of 64 bytes as defined in Layout.md. Consider using ZeroPadding afterwards, to conform to the Arrow layout specification.

Parameters
  • new_size: The new size for the buffer.

  • shrink_to_fit: Whether to shrink the capacity if new size < current size

Status Reserve(const int64_t new_capacity) = 0

Ensure that buffer has enough memory allocated to fit the indicated capacity (and meets the 64 byte padding requirement in Layout.md).

It does not change buffer’s reported size and doesn’t zero the padding.

Memory Pools

MemoryPool *arrow::default_memory_pool()

Return the process-wide default memory pool.

Status arrow::jemalloc_memory_pool(MemoryPool **out)

Return a process-wide memory pool based on jemalloc.

May return NotImplemented if jemalloc is not available.

Status arrow::mimalloc_memory_pool(MemoryPool **out)

Return a process-wide memory pool based on mimalloc.

May return NotImplemented if mimalloc is not available.

MemoryPool *arrow::system_memory_pool()

Return a process-wide memory pool based on the system allocator.

class arrow::MemoryPool

Base class for memory allocation on the CPU.

Besides tracking the number of allocated bytes, the allocator also should take care of the required 64-byte alignment.

Subclassed by arrow::LoggingMemoryPool, arrow::ProxyMemoryPool, arrow::stl::STLMemoryPool< Allocator >

Public Functions

Status Allocate(int64_t size, uint8_t **out) = 0

Allocate a new memory region of at least size bytes.

The allocated region shall be 64-byte aligned.

Status Reallocate(int64_t old_size, int64_t new_size, uint8_t **ptr) = 0

Resize an already allocated memory section.

As by default most default allocators on a platform don’t support aligned reallocation, this function can involve a copy of the underlying data.

void Free(uint8_t *buffer, int64_t size) = 0

Free an allocated region.

Parameters
  • buffer: Pointer to the start of the allocated memory region

  • size: Allocated size located at buffer. An allocator implementation may use this for tracking the amount of allocated bytes as well as for faster deallocation if supported by its backend.

int64_t bytes_allocated() const = 0

The number of bytes that were allocated and not yet free’d through this allocator.

int64_t max_memory() const

Return peak memory allocation in this memory pool.

Return

Maximum bytes allocated. If not known (or not implemented), returns -1

std::string backend_name() const = 0

The name of the backend used by this MemoryPool (e.g. “system” or “jemalloc”).

Public Static Functions

std::unique_ptr<MemoryPool> CreateDefault()

EXPERIMENTAL. Create a new instance of the default MemoryPool.

class arrow::LoggingMemoryPool : public arrow::MemoryPool

Public Functions

Status Allocate(int64_t size, uint8_t **out) override

Allocate a new memory region of at least size bytes.

The allocated region shall be 64-byte aligned.

Status Reallocate(int64_t old_size, int64_t new_size, uint8_t **ptr) override

Resize an already allocated memory section.

As by default most default allocators on a platform don’t support aligned reallocation, this function can involve a copy of the underlying data.

void Free(uint8_t *buffer, int64_t size) override

Free an allocated region.

Parameters
  • buffer: Pointer to the start of the allocated memory region

  • size: Allocated size located at buffer. An allocator implementation may use this for tracking the amount of allocated bytes as well as for faster deallocation if supported by its backend.

int64_t bytes_allocated() const override

The number of bytes that were allocated and not yet free’d through this allocator.

int64_t max_memory() const override

Return peak memory allocation in this memory pool.

Return

Maximum bytes allocated. If not known (or not implemented), returns -1

std::string backend_name() const override

The name of the backend used by this MemoryPool (e.g. “system” or “jemalloc”).

class arrow::ProxyMemoryPool : public arrow::MemoryPool

Derived class for memory allocation.

Tracks the number of bytes and maximum memory allocated through its direct calls. Actual allocation is delegated to MemoryPool class.

Public Functions

Status Allocate(int64_t size, uint8_t **out) override

Allocate a new memory region of at least size bytes.

The allocated region shall be 64-byte aligned.

Status Reallocate(int64_t old_size, int64_t new_size, uint8_t **ptr) override

Resize an already allocated memory section.

As by default most default allocators on a platform don’t support aligned reallocation, this function can involve a copy of the underlying data.

void Free(uint8_t *buffer, int64_t size) override

Free an allocated region.

Parameters
  • buffer: Pointer to the start of the allocated memory region

  • size: Allocated size located at buffer. An allocator implementation may use this for tracking the amount of allocated bytes as well as for faster deallocation if supported by its backend.

int64_t bytes_allocated() const override

The number of bytes that were allocated and not yet free’d through this allocator.

int64_t max_memory() const override

Return peak memory allocation in this memory pool.

Return

Maximum bytes allocated. If not known (or not implemented), returns -1

std::string backend_name() const override

The name of the backend used by this MemoryPool (e.g. “system” or “jemalloc”).

Allocation Functions

These functions allocate a buffer from a particular memory pool.

Result<std::unique_ptr<Buffer>> AllocateBuffer(const int64_t size, MemoryPool *pool = NULLPTR)

Allocate a fixed size mutable buffer from a memory pool, zero its padding.

Parameters
  • [in] size: size of buffer to allocate

  • [in] pool: a memory pool

Result<std::unique_ptr<ResizableBuffer>> AllocateResizableBuffer(const int64_t size, MemoryPool *pool = NULLPTR)

Allocate a resizeable buffer from a memory pool, zero its padding.

Parameters
  • [in] size: size of buffer to allocate

  • [in] pool: a memory pool

Result<std::shared_ptr<Buffer>> AllocateBitmap(int64_t length, MemoryPool *pool = NULLPTR)

Allocate a bitmap buffer from a memory pool no guarantee on values is provided.

Parameters
  • [in] length: size in bits of bitmap to allocate

  • [in] pool: memory pool to allocate memory from

Status AllocateBitmap(MemoryPool *pool, int64_t length, std::shared_ptr<Buffer> *out)
Result<std::shared_ptr<Buffer>> AllocateEmptyBitmap(int64_t length, MemoryPool *pool = NULLPTR)

Allocate a zero-initialized bitmap buffer from a memory pool.

Parameters
  • [in] length: size in bits of bitmap to allocate

  • [in] pool: memory pool to allocate memory from

Result<std::shared_ptr<Buffer>> ConcatenateBuffers(const BufferVector &buffers, MemoryPool *pool = NULLPTR)

Concatenate multiple buffers into a single buffer.

Parameters
  • [in] buffers: to be concatenated

  • [in] pool: memory pool to allocate the new buffer from

Status ConcatenateBuffers(const BufferVector &buffers, MemoryPool *pool, std::shared_ptr<Buffer> *out)

Slicing

std::shared_ptr<Buffer> SliceBuffer(const std::shared_ptr<Buffer> &buffer, const int64_t offset, const int64_t length)

Construct a view on a buffer at the given offset and length.

This function cannot fail and does not check for errors (except in debug builds)

std::shared_ptr<Buffer> SliceBuffer(const std::shared_ptr<Buffer> &buffer, const int64_t offset)

Construct a view on a buffer at the given offset, up to the buffer’s end.

This function cannot fail and does not check for errors (except in debug builds)

Result<std::shared_ptr<Buffer>> SliceBufferSafe(const std::shared_ptr<Buffer> &buffer, int64_t offset)

Input-checking version of SliceBuffer.

An Invalid Status is returned if the requested slice falls out of bounds.

Result<std::shared_ptr<Buffer>> SliceBufferSafe(const std::shared_ptr<Buffer> &buffer, int64_t offset, int64_t length)

Input-checking version of SliceBuffer.

An Invalid Status is returned if the requested slice falls out of bounds. Note that unlike SliceBuffer, length isn’t clamped to the available buffer size.

std::shared_ptr<Buffer> SliceMutableBuffer(const std::shared_ptr<Buffer> &buffer, const int64_t offset, const int64_t length)

Like SliceBuffer, but construct a mutable buffer slice.

If the parent buffer is not mutable, behavior is undefined (it may abort in debug builds).

std::shared_ptr<Buffer> SliceMutableBuffer(const std::shared_ptr<Buffer> &buffer, const int64_t offset)

Like SliceBuffer, but construct a mutable buffer slice.

If the parent buffer is not mutable, behavior is undefined (it may abort in debug builds).

Result<std::shared_ptr<Buffer>> SliceMutableBufferSafe(const std::shared_ptr<Buffer> &buffer, int64_t offset)

Input-checking version of SliceMutableBuffer.

An Invalid Status is returned if the requested slice falls out of bounds.

Result<std::shared_ptr<Buffer>> SliceMutableBufferSafe(const std::shared_ptr<Buffer> &buffer, int64_t offset, int64_t length)

Input-checking version of SliceMutableBuffer.

An Invalid Status is returned if the requested slice falls out of bounds. Note that unlike SliceBuffer, length isn’t clamped to the available buffer size.

Buffer Builders

class arrow::BufferBuilder

A class for incrementally building a contiguous chunk of in-memory data.

Public Functions

BufferBuilder(std::shared_ptr<ResizableBuffer> buffer, MemoryPool *pool = default_memory_pool())

Constructs new Builder that will start using the provided buffer until Finish/Reset are called.

The buffer is not resized.

Status Resize(const int64_t new_capacity, bool shrink_to_fit = true)

Resize the buffer to the nearest multiple of 64 bytes.

Return

Status

Parameters
  • new_capacity: the new capacity of the of the builder. Will be rounded up to a multiple of 64 bytes for padding

  • shrink_to_fit: if new capacity is smaller than the existing size, reallocate internal buffer. Set to false to avoid reallocations when shrinking the builder.

Status Reserve(const int64_t additional_bytes)

Ensure that builder can accommodate the additional number of bytes without the need to perform allocations.

Return

Status

Parameters
  • [in] additional_bytes: number of additional bytes to make space for

Status Append(const void *data, const int64_t length)

Append the given data to the buffer.

The buffer is automatically expanded if necessary.

Status Append(const int64_t num_copies, uint8_t value)

Append copies of a value to the buffer.

The buffer is automatically expanded if necessary.

Status Finish(std::shared_ptr<Buffer> *out, bool shrink_to_fit = true)

Return result of builder as a Buffer object.

The builder is reset and can be reused afterwards.

Return

Status

Parameters
  • [out] out: the finalized Buffer object

  • shrink_to_fit: if the buffer size is smaller than its capacity, reallocate to fit more tightly in memory. Set to false to avoid a reallocation, at the expense of potentially more memory consumption.

void Rewind(int64_t position)

Set size to a smaller value without modifying builder contents.

For reusable BufferBuilder classes

Parameters
  • [in] position: must be non-negative and less than or equal to the current length()

Public Static Functions

int64_t GrowByFactor(int64_t current_capacity, int64_t new_capacity)

Return a capacity expanded by the desired growth factor.

template<typename T, typename Enable = void>
class TypedBufferBuilder

STL Integration

template<class T>
class arrow::stl::allocator

A STL allocator delegating allocations to a Arrow MemoryPool.

Public Functions

allocator() noexcept

Construct an allocator from the default MemoryPool.

allocator(MemoryPool *pool) noexcept

Construct an allocator from the given MemoryPool.

template<class U>
struct rebind
template<typename Allocator = std::allocator<uint8_t>>
class arrow::stl::STLMemoryPool : public arrow::MemoryPool

A MemoryPool implementation delegating allocations to a STL allocator.

Note that STL allocators don’t provide a resizing operation, and therefore any buffer resizes will do a full reallocation and copy.

Public Functions

STLMemoryPool(const Allocator &alloc)

Construct a memory pool from the given allocator.

Status Allocate(int64_t size, uint8_t **out) override

Allocate a new memory region of at least size bytes.

The allocated region shall be 64-byte aligned.

Status Reallocate(int64_t old_size, int64_t new_size, uint8_t **ptr) override

Resize an already allocated memory section.

As by default most default allocators on a platform don’t support aligned reallocation, this function can involve a copy of the underlying data.

void Free(uint8_t *buffer, int64_t size) override

Free an allocated region.

Parameters
  • buffer: Pointer to the start of the allocated memory region

  • size: Allocated size located at buffer. An allocator implementation may use this for tracking the amount of allocated bytes as well as for faster deallocation if supported by its backend.

int64_t bytes_allocated() const override

The number of bytes that were allocated and not yet free’d through this allocator.

int64_t max_memory() const override

Return peak memory allocation in this memory pool.

Return

Maximum bytes allocated. If not known (or not implemented), returns -1

std::string backend_name() const override

The name of the backend used by this MemoryPool (e.g. “system” or “jemalloc”).