Component encapsulating a UCP worker. More...
#include <worker.h>
Classes | |
| struct | Attributes |
| Worker attributes reported by ucp_worker_query. More... | |
Public Member Functions | |
| Worker (const Worker &)=delete | |
| Worker & | operator= (Worker const &)=delete |
| Worker (Worker &&o)=delete | |
| Worker & | operator= (Worker &&o)=delete |
| virtual | ~Worker () |
| ucxx::Worker destructor. | |
| ucp_worker_h | getHandle () |
| Get the underlying ucp_worker_h handle. | |
| std::string | getInfo () |
| Get information about the underlying ucp_worker_h object. | |
| void | initBlockingProgressMode () |
| Initialize blocking progress mode. | |
| int | getEpollFileDescriptor () |
| Get the epoll file descriptor associated with the worker. | |
| bool | arm () |
| Arm the UCP worker. | |
| bool | progressWorkerEvent (const int epollTimeout=-1) |
| Progress worker event while in blocking progress mode. | |
| void | signal () |
| Signal the worker that an event happened. | |
| bool | waitProgress () |
| Block until an event has happened, then progresses. | |
| bool | progressOnce () |
| Progress the worker only once. | |
| bool | progress () |
| Progress the worker until all communication events are completed. | |
| void | registerDelayedSubmission (std::shared_ptr< Request > request, DelayedSubmissionCallbackType callback) |
| Register delayed request submission. | |
| bool | registerGenericPre (DelayedSubmissionCallbackType callback, uint64_t period=0) |
| Register callback to be executed in progress thread before progressing. | |
| bool | registerGenericPost (DelayedSubmissionCallbackType callback, uint64_t period=0) |
| Register callback to be executed in progress thread before progressing. | |
| bool | isDelayedRequestSubmissionEnabled () const |
| Inquire if worker has been created with delayed submission enabled. | |
| bool | isFutureEnabled () const |
| Inquire if worker has been created with future support. | |
| bool | isRequestAttributesEnabled () const noexcept |
| Inquire if worker has been created with request attributes querying enabled. | |
| BufferType | getCudaBufferType () const |
| Get the preferred buffer type for CUDA allocations. | |
| virtual void | populateFuturesPool () |
| Populate the futures pool. | |
| virtual void | clearFuturesPool () |
| Clear the futures pool. | |
| virtual std::shared_ptr< Future > | getFuture () |
| Get a future from the pool. | |
| virtual RequestNotifierWaitState | waitRequestNotifier (uint64_t periodNs) |
| Block until a request event. | |
| virtual void | runRequestNotifier () |
| Notify futures of each completed communication request. | |
| virtual void | stopRequestNotifierThread () |
| Signal the notifier to terminate. | |
| void | setProgressThreadStartCallback (std::function< void(void *)> callback, void *callbackArg) |
| Set callback to be executed at the progress thread start. | |
| void | startProgressThread (const bool pollingMode=false, const int epollTimeout=1) |
| Start the progress thread. | |
| void | stopProgressThread () |
| Stop the progress thread. | |
| bool | isProgressThreadRunning () |
| Inquire if worker has a progress thread running. | |
| std::thread::id | getProgressThreadId () |
| Get the progress thread ID. | |
| size_t | cancelInflightRequests (uint64_t period=0, uint64_t maxAttempts=1) |
| Cancel inflight requests. | |
| void | scheduleRequestCancel (TrackedRequests trackedRequests) |
| Schedule cancelation of inflight requests. | |
| void | removeInflightRequest (std::shared_ptr< Request > request) |
| Remove reference to request from internal container. | |
| std::shared_ptr< TagProbeInfo > | tagProbe (const Tag tag, const TagMask tagMask=TagMaskFull, const bool remove=false) const |
| Check for uncaught tag messages. | |
| RequestTagBuilder | tagRecvBuilder (void *buffer, size_t length, Tag tag, TagMask tagMask) |
| Create a builder for a tag receive operation. | |
| RequestTagBuilder | tagRecvWithHandleBuilder (void *buffer, std::shared_ptr< TagProbeInfo > probeInfo) |
| Create a builder for a tag receive operation using a message handle. | |
| RequestTagBuilder | tagRecvWithHandleBuilder (void *buffer, size_t length, std::shared_ptr< TagProbeInfo > probeInfo) |
| Create a builder for a tag receive operation using a message handle. | |
| AddressBuilder | addressBuilder () |
| Create a builder for this worker's address. | |
| EndpointBuilder | endpointBuilder (std::string ipAddress, uint16_t port) |
| Create a builder for an endpoint to a worker listening on an IP and port. | |
| EndpointBuilder | endpointBuilder (std::shared_ptr< Address > address) |
| Create a builder for an endpoint to a worker located at a UCX address. | |
| ListenerBuilder | listenerBuilder (uint16_t port, ucp_listener_conn_callback_t callback, void *callbackArgs) |
| Create a builder for a listener on this worker. | |
| void | registerAmAllocator (ucs_memory_type_t memoryType, AmAllocatorType allocator) |
| Register allocator for active messages. | |
| void | registerAmReceiverCallback (AmReceiverCallbackInfo info, AmReceiverCallbackType callback) |
| Register receiver callback for active messages. | |
| bool | amProbe (const ucp_ep_h endpointHandle) const |
| Check for uncaught active messages. | |
| RequestFlushBuilder | flushBuilder () |
| Create a builder for a flush operation. | |
| Attributes | queryAttributes () const |
| Get the worker's attributes. | |
| Public Member Functions inherited from ucxx::Component | |
| void | setParent (std::shared_ptr< Component > parent) |
| Set the internal parent reference. | |
| std::shared_ptr< Component > | getParent () const |
| Get the internal parent reference. | |
Protected Member Functions | |
| Worker (std::shared_ptr< Context > context, const bool enableDelayedSubmission=false, const bool enableFuture=false) | |
| Protected constructor of ucxx::Worker. | |
Protected Attributes | |
| bool | _enableFuture |
| Boolean identifying whether the worker was created with future capability. | |
| bool | _enableRequestAttributes |
| Whether request attributes (e.g. UCP debug info) are queried for each request. | |
| std::mutex | _futuresPoolMutex {} |
| Mutex to access the futures pool. | |
| std::queue< std::shared_ptr< Future > > | _futuresPool {} |
| Futures pool to prevent running out of fresh futures. | |
| std::shared_ptr< Notifier > | _notifier {nullptr} |
| Notifier object. | |
| std::shared_ptr< internal::AmData > | _amData |
| Worker data made available to Active Messages callback. | |
| BufferType | _cudaBufferType {BufferType::Invalid} |
| Preferred buffer type for CUDA allocations. | |
| Protected Attributes inherited from ucxx::Component | |
| std::shared_ptr< Component > | _parent {nullptr} |
| A reference-counted pointer to the parent. | |
Friends | |
| class | detail::ConstructorFactory |
| class | WorkerBuilder |
| Allow WorkerBuilder to access protected/private constructor. | |
| void | detail::registerInflightRequest (std::shared_ptr< Component > component, std::shared_ptr< Request > req) |
| Allow request builders to register newly-created requests. | |
Component encapsulating a UCP worker.
The UCP layer provides a handle to access workers in form of ucp_worker_h object, this class encapsulates that object and provides methods to simplify its handling.
|
explicitprotected |
Protected constructor of ucxx::Worker.
This is the internal implementation of ucxx::Worker constructor, made protected not to be called directly.
Instead the user should use one of the following:
| [in] | context | the context from which to create the worker. |
| [in] | enableDelayedSubmission | if true, each ucxx::Request will not be submitted immediately, but instead delayed to the progress thread. Requires use of the progress thread. |
| [in] | enableFuture | if true, notifies the future associated with each ucxx::Request, currently used only by ucxx::python::Worker. |
|
nodiscard |
Create a builder for this worker's address.
Calling this method only creates the builder. Finalizing it with .build() or implicit conversion gets the address of the underlying UCX worker object.
|
nodiscard |
Check for uncaught active messages.
Checks the worker for any uncaught active messages. An uncaught active message is any active message that has been fully or partially received by the worker, but not matched by a corresponding active-message receive request.
| bool ucxx::Worker::arm | ( | ) |
Arm the UCP worker.
Wrapper for ucp_worker_arm, checking its return status for errors and raising an exception if an error occurred.
| ucxx::Error | if an error occurred while attempting to arm the worker. |
| size_t ucxx::Worker::cancelInflightRequests | ( | uint64_t | period = 0, |
| uint64_t | maxAttempts = 1 ) |
Cancel inflight requests.
Cancel inflight requests, returning the total number of requests that were canceled. This is usually executed during the progress loop.
If the parent worker is running a progress thread, a maximum timeout may be specified for which the close operation will wait. This can be particularly important for cases where the progress thread might be attempting to acquire a resource (e.g., the Python GIL) while the current thread owns that resource. In particular for Python, the ~Worker() will call this method for which we can't release the GIL when the garbage collector runs and destroys the object.
| [in] | period | maximum period to wait for a generic pre/post progress thread operation will wait for. |
| [in] | maxAttempts | maximum number of attempts to close endpoint, only applicable if worker is running a progress thread and period > 0. |
|
virtual |
Clear the futures pool.
Clear the futures pool, ensuring all references are removed and thus avoiding reference cycles that prevent the ucxx::Worker and other resources from cleaning up on time.
| std::runtime_error | if future support is not implemented. |
Reimplemented in ucxx::python::Worker.
|
nodiscard |
Create a builder for an endpoint to a worker located at a UCX address.
Calling this method only creates the builder. Finalizing it with .build() or implicit conversion creates the endpoint from this worker.
| [in] | address | address of the remote UCX worker. |
|
nodiscard |
Create a builder for an endpoint to a worker listening on an IP and port.
Calling this method only creates the builder. Finalizing it with .build() or implicit conversion creates the endpoint from this worker.
| [in] | ipAddress | string containing the IP address of the remote worker. |
| [in] | port | port number where the remote worker is listening at. |
|
nodiscard |
Create a builder for a flush operation.
Calling this method only creates the builder. Finalizing it with .build() or implicit conversion submits the request.
|
nodiscard |
Get the preferred buffer type for CUDA allocations.
Returns the buffer type used when allocating CUDA buffers for incoming multi-buffer tag receives. Defaults to CCCL if compiled with CCCL support, otherwise Invalid.
| int ucxx::Worker::getEpollFileDescriptor | ( | ) |
Get the epoll file descriptor associated with the worker.
Get the epoll file descriptor associated with the worker when running in blocking mode. The worker only has an associated epoll file descriptor after initBlockingProgressMode() is executed.
The file descriptor is destroyed as part of the ucxx::Worker destructor, thus any reference to it shall not be used after that.
| std::runtime_error | if initBlockingProgressMode() was not executed to run the worker in blocking progress mode. |
|
nodiscardvirtual |
Get a future from the pool.
Get a future from the pool. If the pool is empty, ucxx::Worker::populateFuturesPool() is called and a warning is raised, since that likely means the user is missing to call the aforementioned method regularly.
| std::runtime_error | if future support is not implemented. |
Reimplemented in ucxx::python::Worker.
|
nodiscard |
Get the underlying ucp_worker_h handle.
Lifetime of the ucp_worker_h handle is managed by the ucxx::Worker object and its ownership is non-transferrable. Once the ucxx::Worker is destroyed the handle is not valid anymore, it is the user's responsibility to ensure the owner's lifetime while using the handle.
|
nodiscard |
Get information about the underlying ucp_worker_h object.
Convenience wrapper for ucp_worker_print_info() to get information about the underlying UCP worker handle and return it as a string.
|
nodiscard |
Get the progress thread ID.
Get the progress thread ID, only valid if startProgressThread() was called.
| void ucxx::Worker::initBlockingProgressMode | ( | ) |
Initialize blocking progress mode.
Initialize blocking progress mode, creates internal file descriptors to handle blocking progress by waiting for the UCP worker to notify the file descriptors. This method is supposed to be called when usage of progressWorkerEvent() is intended, before the first call to progressWorkerEvent(). If using polling mode only via progress()/progressOnce() calls or wake-up with waitProgress(), this method should not be called.
In blocking mode, the user should call progressWorkerEvent() to block and then progress the worker as new events arrive. wakeProgressEvent() may be called to forcefully wake this method, for example to shutdown the application.
| std::ios_base::failure | if creating any of the file descriptors or setting their statuses. |
|
nodiscard |
Inquire if worker has been created with delayed submission enabled.
Check whether the worker has been created with delayed submission enabled.
|
nodiscard |
Inquire if worker has been created with future support.
Check whether the worker has been created with future support.
|
nodiscard |
Inquire if worker has a progress thread running.
Check whether the worker currently has a progress thread running.
|
nodiscardnoexcept |
Inquire if worker has been created with request attributes querying enabled.
Check whether the worker has been created with request attributes querying enabled. When enabled, each ucxx::Request will have its UCP attributes (such as the debug string) queried immediately after submission, making them available via ucxx::Request::queryAttributes(). Querying request attributes has a non-negligible runtime cost and is therefore disabled by default.
|
nodiscard |
Create a builder for a listener on this worker.
Calling this method only creates the builder. Finalizing it with .build() or implicit conversion starts a listener on this worker.
| [in] | port | port number where to listen at. |
| [in] | callback | callback to handle each incoming connection. |
| [in] | callbackArgs | pointer to argument to pass to the callback. |
|
virtual |
Populate the futures pool.
To avoid taking blocking resources (such as the Python GIL) for every new future required by each ucxx::Request, the ucxx::Worker maintains a pool of futures that can be acquired when a new ucxx::Request is created. Currently the pool has a maximum size of 100 objects, and will refill once it goes under 50, otherwise calling this functions results in a no-op.
| std::runtime_error | if future support is not implemented. |
Reimplemented in ucxx::python::Worker.
| bool ucxx::Worker::progress | ( | ) |
Progress the worker until all communication events are completed.
Iteratively calls progressOnce() until all communication events are completed. Additionally ensure inflight messages pending for cancelation are canceled.
| bool ucxx::Worker::progressOnce | ( | ) |
Progress the worker only once.
Wrapper for ucp_worker_progress.
| bool ucxx::Worker::progressWorkerEvent | ( | const int | epollTimeout = -1 | ) |
Progress worker event while in blocking progress mode.
Blocks until a new worker event has happened and the worker notifies the file descriptor associated with it, or epollTimeout has elapsed. Requires blocking progress mode to be initialized with initBlockingProgressMode() before the first call to this method. Additionally ensure inflight messages pending for cancelation are canceled.
| [in] | epollTimeout | timeout in ms when waiting for worker event, or -1 to block indefinitely. |
| std::ios_base::failure | if creating any of the file descriptors or setting their statuses. |
|
nodiscard |
Get the worker's attributes.
Returns the worker attributes as a struct, querying UCP via ucp_worker_query under the hood. All non-address fields exposed by UCP are queried, use addressBuilder() to obtain the address.
| ucxx::Error | if an error occurred while querying worker attributes. |
| void ucxx::Worker::registerAmAllocator | ( | ucs_memory_type_t | memoryType, |
| AmAllocatorType | allocator ) |
Register allocator for active messages.
Register a new allocator for active messages. By default, only one allocator is defined for host memory (UCS_MEMORY_TYPE_HOST), and is used as a fallback when an allocator for the source's memory type is unavailable. In many circumstances relying exclusively on the host allocator is undesirable, for example when transferring CUDA buffers the destination is always going to be a host buffer and prevent the use of transports such as NVLink or InfiniBand+GPUDirectRDMA. For that reason it's important that the user defines those allocators that are important for the application.
If the memoryType has already been registered, the previous allocator will be replaced by the new one. Be careful when doing this after transfers have started, there are no guarantees that inflight messages have not already been allocated with the old allocator for that type.
| [in] | memoryType | the memory type the allocator will be used for. |
| [in] | allocator | the allocator callable that will be used to allocate new active message buffers. |
| void ucxx::Worker::registerAmReceiverCallback | ( | AmReceiverCallbackInfo | info, |
| AmReceiverCallbackType | callback ) |
Register receiver callback for active messages.
Register a new receiver callback for active messages. By default, active messages do not execute any callbacks on the receiving end unless one is specified when sending the message. If the message sender specifies a callback receiver identifier then the remote receiver needs to have a callback registered with the same identifier to execute when the request completes. To ensure multiple applications that do not know about each other can have coexisting callbacks where receiver identifiers may have the same value, an owner must be specified as well, which has the form of a string and should be reasonably unique to prevent accidentally calling callbacks from a separate application, thus names like "A" or "UCX" are discouraged in favor of more descriptive names such as "MyFastCommsProject", and the name "ucxx" is reserved.
Because it is impossible to predict which callback would be called in such an event, the registered callback cannot be changed, thus calling this method with the same given owner and identifier will throw std::runtime_error.
| std::runtime_error | if a callback with same given owner and identifier is already registered, or if the reserved owner name "ucxx" is specified. |
| [in] | info | the owner name and unique identifier of the receiver callback. |
| [in] | callback | the callback to execute when the active message is received. |
| void ucxx::Worker::registerDelayedSubmission | ( | std::shared_ptr< Request > | request, |
| DelayedSubmissionCallbackType | callback ) |
Register delayed request submission.
Register ucxx::Request for delayed submission. When the ucxx::Worker is created with enableDelayedSubmission=true, calling actual UCX transfer routines will not happen immediately and instead will be submitted later by the worker thread.
The purpose of this method is to offload as much as possible any work to the worker thread, thus decreasing computation on the caller thread, but potentially increasing transfer latency.
| [in] | request | the request to which the callback belongs, ensuring it remains alive until the callback is invoked. |
| [in] | callback | the callback set to execute the UCP transfer routine during the worker thread loop. |
|
nodiscard |
Register callback to be executed in progress thread before progressing.
Register callback to be executed in the current or next iteration of the progress thread after the worker is progressed. There is no guarantee that the callback will be executed in the current or next iteration, this depends on where the progress thread is in the current iteration when this callback is registered.
The purpose of this method is to schedule operations to be executed in the progress thread, immediately after progressing the worker completes.
If period is 0 this is a blocking call that only returns when the callback has been executed and will always return true, and if period is a positive integer the time in nanoseconds will be waited for the callback to complete and return true in the successful case or false otherwise. However, if the callback is not cancelable anymore (i.e., it has already started), this method will keep retrying and may never return if the callback never completes, it is unsafe to return as this would allow the caller to destroy the callback and its resources causing undefined behavior. period only applies if the worker progress thread is running, otherwise the callback is immediately executed.
| [in] | callback | the callback to execute before progressing the worker. |
| [in] | period | the time in nanoseconds to wait for the callback to complete. |
|
nodiscard |
Register callback to be executed in progress thread before progressing.
Register callback to be executed in the current or next iteration of the progress thread before the worker is progressed. There is no guarantee that the callback will be executed in the current or next iteration, this depends on where the progress thread is in the current iteration when this callback is registered.
The purpose of this method is to schedule operations to be executed in the progress thread, such as endpoint creation and closing, so that progressing doesn't ever need to occur in the application thread when using a progress thread.
If period is 0 this is a blocking call that only returns when the callback has been executed and will always return true, and if period is a positive integer the time in nanoseconds will be waited for the callback to complete and return true in the successful case or false otherwise. However, if the callback is not cancelable anymore (i.e., it has already started), this method will keep retrying and may never return if the callback never completes, it is unsafe to return as this would allow the caller to destroy the callback and its resources causing undefined behavior. period only applies if the worker progress thread is running, otherwise the callback is immediately executed.
| [in] | callback | the callback to execute before progressing the worker. |
| [in] | period | the time in nanoseconds to wait for the callback to complete. |
| void ucxx::Worker::removeInflightRequest | ( | std::shared_ptr< Request > | request | ) |
Remove reference to request from internal container.
Remove the reference to a specific request from the internal container. This should be called when a request has completed and the ucxx::Worker does not need to keep track of it anymore.
| [in] | request | shared pointer to the request |
|
virtual |
Notify futures of each completed communication request.
Notifies futures of each completed communication request of their new status. This method is intended to be used from the Notifier (such as the Python thread running it), where the thread will call waitRequestNotifier() and block until some communication is completed, and then call this method to notify all futures. If this is notifying a Python future, the thread where this method is called from must be using the same Python event loop as the thread that submitted the transfer request.
| std::runtime_error | if future support is not implemented. |
Reimplemented in ucxx::python::Worker.
| void ucxx::Worker::scheduleRequestCancel | ( | TrackedRequests | trackedRequests | ) |
Schedule cancelation of inflight requests.
Schedule inflight request to be canceled when cancelInflightRequests() is executed the next time, usually during the progress loop. This is usually called from a ucxx::Endpoint, for example when the error callback was called, signaling that inflight requests for that endpoint will not be completed successfully and should be canceled.
| [in] | trackedRequests | the requests tracked by a child of this class to be scheduled for cancelation. |
| void ucxx::Worker::setProgressThreadStartCallback | ( | std::function< void(void *)> | callback, |
| void * | callbackArg ) |
Set callback to be executed at the progress thread start.
Sets a callback that will be executed at the beginning of the progress thread. This can be used to initialize any resources that are required to be available on the thread the worker will be progressed from, such as a CUDA context.
| [in] | callback | function to execute during progress thread start |
| [in] | callbackArg | argument to be passed to the callback function |
| void ucxx::Worker::signal | ( | ) |
Signal the worker that an event happened.
Signals that an event has happened while, causing both either progressWorkerEvent() or waitProgress() to immediately wake-up.
| ucxx::Error | if an error occurred while attempting to signal the worker. |
| void ucxx::Worker::startProgressThread | ( | const bool | pollingMode = false, |
| const int | epollTimeout = 1 ) |
Start the progress thread.
Spawns a new thread that will take care of continuously progressing the worker. The thread can progress the worker in blocking mode, using progressWorkerEvent() only when worker events happen, or in polling mode by continuously calling progress() (incurs in high CPU utilization).
| [in] | pollingMode | use polling mode if true, or blocking mode if false. |
| [in] | epollTimeout | timeout in ms when waiting for worker event, or -1 to block indefinitely, only applicable if pollingMode==true. |
| void ucxx::Worker::stopProgressThread | ( | ) |
Stop the progress thread.
Stop the progress thread.
May be called by the user at any time, and also called during destructor if the worker thread was ever started.
|
virtual |
Signal the notifier to terminate.
Signals the notifier to terminate, awakening the waitRequestNotifier() blocking call.
| std::runtime_error | if future support is not implemented. |
Reimplemented in ucxx::python::Worker.
|
nodiscard |
Check for uncaught tag messages.
Checks the worker for any uncaught tag messages. An uncaught tag message is any tag message that has been fully or partially received by the worker, but not matched by a corresponding ucp_tag_recv_* call. Additionally, returns information about the tag message.
When remove is false (default), the message remains in the receive queue and a normal tag receive operation must be posted to consume it. When remove is true, the message is removed from the queue and a message handle is returned that can be passed directly to tagRecvWithHandleBuilder for efficient consumption, in this case the caller is required to build a tagRecvWithHandleBuilder request, otherwise a warning will be thrown during destruction of the returned object.
Note this is a non-blocking call, if this is being used to actively check for an incoming message the worker should be constantly progress until a valid probe is returned.
| [in] | tag | the tag to match. |
| [in] | tagMask | the tag mask to use. |
| [in] | remove | if true, remove the message from the queue and return a message handle; if false, leave the message in the queue. |
|
nodiscard |
Create a builder for a tag receive operation.
Calling this method only creates the builder. Finalizing it with .build() or implicit conversion submits the request.
| [in] | buffer | a raw pointer to pre-allocated memory where resulting data will be stored. |
| [in] | length | the size in bytes of the message to be received. |
| [in] | tag | the tag to match. |
| [in] | tagMask | the tag mask to use. |
|
nodiscard |
Create a builder for a tag receive operation using a message handle.
If the message does not fit in the receive buffer, the resulting request completes with UCS_ERR_MESSAGE_TRUNCATED. Calling Request::checkError() then raises ucxx::MessageTruncatedError.
| [in] | buffer | a raw pointer to pre-allocated receive memory. |
| [in] | length | the size in bytes of the receive buffer. |
| [in] | probeInfo | the TagProbeInfo object containing message length and handle. |
|
nodiscard |
Create a builder for a tag receive operation using a message handle.
Calling this method only creates the builder. Finalizing it with .build() or implicit conversion submits the request.
| [in] | buffer | a raw pointer to pre-allocated memory where resulting data will be stored. |
| [in] | probeInfo | the TagProbeInfo object containing message length and handle. |
| bool ucxx::Worker::waitProgress | ( | ) |
Block until an event has happened, then progresses.
Blocks until an event has happened as part of UCX's wake-up mechanism and progress the worker. Additionally ensure inflight messages pending for cancelation are canceled.
| ucxx::Error | if an error occurred while attempting to arm the worker. |
|
nodiscardvirtual |
Block until a request event.
Blocks until some communication is completed and future is ready to be notified, shutdown was initiated or a timeout occurred (only if periodNs > 0). This method is intended for use from the notifier (such as the Python thread running it), where that thread will block until one of the aforementioned events occur.
| std::runtime_error | if future support is not implemented. |
Reimplemented in ucxx::python::Worker.
|
protected |
Boolean identifying whether the worker was created with future capability.
|
protected |
Whether request attributes (e.g. UCP debug info) are queried for each request.