Attention

The vector search and clustering algorithms in RAFT are being migrated to a new library dedicated to vector search called cuVS. We will continue to support the vector search algorithms in RAFT during this move, but will no longer update them after the RAPIDS 24.06 (June) release. We plan to complete the migration by RAPIDS 24.10 (October) release and they will be removed from RAFT altogether in the 24.12 (December) release.

Interruptible#

#include <raft/core/interruptible.hpp>

namespace raft::core

struct interrupted_exception : public raft::exception#
#include <interruptible.hpp>

Exception thrown during interruptible::synchronize call when it detects a request to cancel the work performed in this CPU thread.

class interruptible#
#include <interruptible.hpp>

Cooperative-style interruptible execution.

This class provides facilities for interrupting execution of a C++ thread at designated points in code from outside of the thread. In particular, it provides an interruptible version of the blocking CUDA synchronization function, that allows dropping a long-running GPU work.

Important: Although CUDA synchronize calls serve as cancellation points, the interruptible machinery has nothing to do with CUDA streams or events. In other words, when you call cancel, it’s the CPU waiting function what is interrupted, not the GPU stream work. This means, when the interrupted_exception is raised, any unfinished GPU stream work continues to run. It’s the responsibility of the developer then to make sure the unfinished stream work does not affect the program in an undesirable way.

What can happen to CUDA stream when the synchronize is cancelled? If you catch the interrupted_exception immediately, you can safely wait on the stream again. Otherwise, some of the allocated resources may be released before the active kernel finishes using them, which will result in writing into deallocated or reallocated memory and undefined behavior in general. A dead-locked kernel may never finish (or may crash if you’re lucky). In practice, the outcome is usually acceptable for the use case of emergency program interruption (e.g., CTRL+C), but extra effort on the use side is required to allow safe interrupting and resuming of the GPU stream work.

Public Functions

inline void cancel() noexcept#

Cancel any current or next call to interruptible::synchronize performed on the CPU thread given by this interruptible token.

Note, this function does not involve thread synchronization/locks and does not throw any exceptions, so it’s safe to call from a signal handler.

Public Static Functions

static inline void synchronize(rmm::cuda_stream_view stream)#

Synchronize the CUDA stream, subject to being interrupted by interruptible::cancel called on this CPU thread.

Parameters:

stream[in] a CUDA stream.

Throws:
static inline void synchronize(cudaEvent_t event)#

Synchronize the CUDA event, subject to being interrupted by interruptible::cancel called on this CPU thread.

Parameters:

event[in] a CUDA event.

Throws:
static inline void yield()#

Check the thread state, whether the thread can continue execution or is interrupted by interruptible::cancel.

This is a cancellation point for an interruptible thread. It’s called in the internals of interruptible::synchronize in a loop. If two synchronize calls are far apart, it’s recommended to call interruptible::yield() in between to make sure the thread does not become unresponsive for too long.

Both yield and yield_no_throw reset the state to non-cancelled after execution.

Throws:

raft::interrupted_exception – if interruptible::cancel() was called on the current CPU thread.

static inline auto yield_no_throw() -> bool#

Check the thread state, whether the thread can continue execution or is interrupted by interruptible::cancel.

Same as interruptible::yield, but does not throw an exception if the thread is cancelled.

Both yield and yield_no_throw reset the state to non-cancelled after execution.

Returns:

whether the thread can continue, i.e. true means continue, false means cancelled.

static inline auto get_token() -> std::shared_ptr<interruptible>#

Get a cancellation token for this CPU thread.

Returns:

an object that can be used to cancel the GPU work waited on this CPU thread.

static inline auto get_token(std::thread::id thread_id) -> std::shared_ptr<interruptible>#

Get a cancellation token for a CPU thread given by its id.

The returned token may live longer than the associated thread. In that case, using its cancel method has no effect.

Parameters:

thread_id[in] an id of a C++ CPU thread.

Returns:

an object that can be used to cancel the GPU work waited on the given CPU thread.

static inline void cancel(std::thread::id thread_id)#

Cancel any current or next call to interruptible::synchronize performed on the CPU thread given by the thread_id

Note, this function uses a mutex to safely get a cancellation token that may be shared among multiple threads. If you plan to use it from a signal handler, consider the non-static cancel() instead.

Parameters:

thread_id[in] a CPU thread, in which the work should be interrupted.