Skip to content

Latest commit

 

History

History
320 lines (239 loc) · 13.5 KB

external-memory.rst

File metadata and controls

320 lines (239 loc) · 13.5 KB
Raw

External Memory Management (EMM) Plugin interface

The CUDA Array Interface <cuda-array-interface> enables sharing of data between different Python libraries that access CUDA devices. However, each library manages its own memory distinctly from the others. For example:

  • By default, Numba allocates memory on CUDA devices by interacting with the CUDA driver API to call functions such as cuMemAlloc and cuMemFree, which is suitable for many use cases.
  • The RAPIDS libraries (cuDF, cuML, etc.) use the RAPIDS Memory Manager (RMM) for allocating device memory.
  • CuPy includes a memory pool implementation for both device and pinned memory.

When multiple CUDA-aware libraries are used together, it may be preferable for Numba to defer to another library for memory management. The EMM Plugin interface facilitates this, by enabling Numba to use another CUDA-aware library for all allocations and deallocations.

An EMM Plugin is used to facilitate the use of an external library for memory management. An EMM Plugin can be a part of an external library, or could be implemented as a separate library.

Overview of External Memory Management

When an EMM Plugin is in use (see setting-emm-plugin), Numba will make memory allocations and deallocations through the Plugin. It will never directly call functions such as cuMemAlloc, cuMemFree, etc.

EMM Plugins always take responsibility for the management of device memory. However, not all CUDA-aware libraries also support managing host memory, so a facility for Numba to continue the management of host memory whilst ceding control of device memory to the EMM is provided (see host-only-cuda-memory-manager).

Effects on Deallocation Strategies

Numba's internal deallocation-behavior is designed to increase efficiency by deferring deallocations until a significant quantity are pending. It also provides a mechanism for preventing deallocations entirely during critical sections, using the ~numba.cuda.defer_cleanup context manager.

When an EMM Plugin is in use, the deallocation strategy is implemented by the EMM, and Numba's internal deallocation mechanism is not used. The EMM Plugin could implement:

  • A similar strategy to the Numba deallocation behaviour, or
  • Something more appropriate to the plugin - for example, deallocated memory might immediately be returned to a memory pool.

The defer_cleanup context manager may behave differently with an EMM Plugin - an EMM Plugin should be accompanied by documentation of the behaviour of the defer_cleanup context manager when it is in use. For example, a pool allocator could always immediately return memory to a pool even when the context manager is in use, but could choose not to free empty pools until defer_cleanup is not in use.

Management of other objects

In addition to memory, Numba manages the allocation and deallocation of events <events>, streams <streams>, and modules (a module is a compiled object, which is generated from @cuda.jit-ted functions). The management of events, streams, and modules is unchanged by the use of an EMM Plugin.

Asynchronous allocation and deallocation

The present EMM Plugin interface does not provide support for asynchronous allocation and deallocation. This may be added to a future version of the interface.

Implementing an EMM Plugin

An EMM Plugin is implemented by deriving from ~numba.cuda.BaseCUDAMemoryManager. A summary of considerations for the implementation follows:

  • Numba instantiates one instance of the EMM Plugin class per context. The context that owns an EMM Plugin object is accessible through self.context, if required.
  • The EMM Plugin is transparent to any code that uses Numba - all its methods are invoked by Numba, and never need to be called by code that uses Numba.
  • The allocation methods memalloc, memhostalloc, and mempin, should use the underlying library to allocate and/or pin device or host memory, and construct an instance of a memory pointer <memory-pointers> representing the memory to return back to Numba. These methods are always called when the current CUDA context is the context that owns the EMM Plugin instance.
  • The initialize method is called by Numba prior to the first use of the EMM Plugin object for a context. This method should do anything required to prepare the underlying library for allocations in the current context. This method may be called multiple times, and must not invalidate previous state when it is called.
  • The reset method is called when all allocations in the context are to be cleaned up. It may be called even prior to initialize, and an EMM Plugin implementation needs to guard against this.
  • To support inter-GPU communication, the get_ipc_handle method should provide an ~numba.cuda.IpcHandle for a given ~numba.cuda.MemoryPointer instance. This method is part of the EMM interface (rather than being handled within Numba) because the base address of the allocation is only known by the underlying library. Closing an IPC handle is handled internally within Numba.
  • It is optional to provide memory info from the get_memory_info method, which provides a count of the total and free memory on the device for the context. It is preferable to implement the method, but this may not be practical for all allocators. If memory info is not provided, this method should raise a RuntimeError.
  • The defer_cleanup method should return a context manager that ensures that expensive cleanup operations are avoided whilst it is active. The nuances of this will vary between plugins, so the plugin documentation should include an explanation of how deferring cleanup affects deallocations, and performance in general.
  • The interface_version property is used to ensure that the plugin version matches the interface provided by the version of Numba. At present, this should always be 1.

Full documentation for the base class follows:

numba.cuda.BaseCUDAMemoryManager

The Host-Only CUDA Memory Manager

Some external memory managers will support management of on-device memory but not host memory. For implementing EMM Plugins using one of these memory managers, a partial implementation of a plugin that implements host-side allocation and pinning is provided. To use it, derive from ~numba.cuda.HostOnlyCUDAMemoryManager instead of ~numba.cuda.BaseCUDAMemoryManager. Guidelines for using this class are:

  • The host-only memory manager implements memhostalloc and mempin - the EMM Plugin should still implement memalloc.
  • If reset is overridden, it must also call super().reset() to allow the host allocations to be cleaned up.
  • If defer_cleanup is overridden, it must hold an active context manager from super().defer_cleanup() to ensure that host-side cleanup is also deferred.

Documentation for the methods of ~numba.cuda.HostOnlyCUDAMemoryManager follows:

numba.cuda.HostOnlyCUDAMemoryManager

The IPC Handle Mixin

An implementation of the get_ipc_handle() function is is provided in the GetIpcHandleMixin class. This uses the driver API to determine the base address of an allocation for opening an IPC handle. If this implementation is appropriate for an EMM plugin, it can be added by mixing in the GetIpcHandleMixin class:

numba.cuda.GetIpcHandleMixin

Classes and structures of returned objects

This section provides an overview of the classes and structures that need to be constructed by an EMM Plugin.

Memory Pointers

EMM Plugins should construct memory pointer instances that represent their allocations, for return to Numba. The appropriate memory pointer class to use in each method is:

  • ~numba.cuda.MemoryPointer: returned from memalloc
  • ~numba.cuda.MappedMemory: returned from memhostalloc or mempin when the host memory is mapped into the device memory space.
  • ~numba.cuda.PinnedMemory: return from memhostalloc or mempin when the host memory is not mapped into the device memory space.

Memory pointers can take a finalizer, which is a function that is called when the buffer is no longer needed. Usually the finalizer will make a call to the memory management library (either internal to Numba, or external if allocated by an EMM Plugin) to inform it that the memory is no longer required, and that it could potentially be freed and/or unpinned. The memory manager may choose to defer actually cleaning up the memory to any later time after the finalizer runs - it is not required to free the buffer immediately.

Documentation for the memory pointer classes follows.

numba.cuda.MemoryPointer

The AutoFreePointer class need not be used directly, but is documented here as it is subclassed by numba.cuda.MappedMemory:

numba.cuda.cudadrv.driver.AutoFreePointer

numba.cuda.MappedMemory

numba.cuda.PinnedMemory

Memory Info

If an implementation of ~numba.cuda.BaseCUDAMemoryManager.get_memory_info is to provide a result, then it should return an instance of the MemoryInfo named tuple:

numba.cuda.MemoryInfo

IPC

An instance of IpcHandle is required to be returned from an implementation of ~numba.cuda.BaseCUDAMemoryManager.get_ipc_handle:

numba.cuda.IpcHandle

Guidance for constructing an IPC handle in the context of implementing an EMM Plugin:

  • The memory parameter passed to the get_ipc_handle method of an EMM Plugin can be passed as the base parameter.
  • A suitable type for the handle can be constructed as ctypes.c_byte * 64. The data for handle must be populated using a method for obtaining a CUDA IPC handle appropriate to the underlying library.
  • size should match the size of the original allocation, which can be obtained with memory.size in get_ipc_handle.
  • An appropriate value for source_info can be created by calling self.context.device.get_device_identity().
  • If the underlying memory does not point to the base of an allocation returned by the CUDA driver or runtime API (e.g. if a pool allocator is in use) then the offset from the base must be provided.

Setting the EMM Plugin

By default, Numba uses its internal memory management - if an EMM Plugin is to be used, it must be configured. There are two mechanisms for configuring the use of an EMM Plugin: an environment variable, and a function.

Environment variable

A module name can be provided in the environment variable, NUMBA_CUDA_MEMORY_MANAGER. If this environment variable is set, Numba will attempt to import the module, and and use its _numba_memory_manager global variable as the memory manager class. This is primarily useful for running the Numba test suite with an EMM Plugin, e.g.:

$ NUMBA_CUDA_MEMORY_MANAGER=rmm python -m numba.runtests numba.cuda.tests

Function

The ~numba.cuda.set_memory_manager function can be used to set the memory manager at runtime. This should be called prior to the initialization of any contexts, as EMM Plugin instances are instantiated along with contexts.

numba.cuda.set_memory_manager

Resetting the memory manager

It is recommended that the memory manager is set once prior to using any CUDA functionality, and left unchanged for the remainder of execution. It is possible to set the memory manager multiple times, noting the following:

  • At the time of their creation, contexts are bound to an instance of a memory manager for their lifetime.
  • Changing the memory manager will have no effect on existing contexts - only contexts created after the memory manager was updated will use instances of the new memory manager.
  • numba.cuda.close can be used to destroy contexts after setting the memory manager so that they get re-created with the new memory manager.
    • This will invalidate any arrays, streams, events, and modules owned by the context.
    • Attempting to use invalid arrays, streams, or events will likely fail with an exception being raised due to a CUDA_ERROR_INVALID_CONTEXT or CUDA_ERROR_CONTEXT_IS_DESTROYED return code from a Driver API function.
    • Attempting to use an invalid module will result in similar, or in some cases a segmentation fault / access violation.

Note

The invalidation of modules means that all functions compiled with @cuda.jit prior to context destruction will need to be redefined, as the code underlying them will also have been unloaded from the GPU.