resiliency

In HPX, a program failure is a manifestation of a failing task. This module exposes several APIs that allow users to manage failing tasks in a convenient way by either replaying a failed task or by replicating a specific task.

Task replay is analogous to the Checkpoint/Restart mechanism found in conventional execution models. The key difference being localized fault detection. When the runtime detects an error, it replays the failing task as opposed to completely rolling back the entire program to the previous checkpoint.

Task replication is designed to provide reliability enhancements by replicating a set of tasks and evaluating their results to determine a consensus among them. This technique is most effective in situations where there are few tasks in the critical path of the DAG which leaves the system underutilized or where hardware or software failures may result in an incorrect result instead of an error. However, the drawback of this method is the additional computational cost incurred by repeating a task multiple times.

The following API functions are exposed:

• hpx::resiliency::experimental::async_replay: This version of task replay will catch user-defined exceptions and automatically reschedule the task N times before throwing an hpx::resiliency::experimental::abort_replay_exception if no task is able to complete execution without an exception.

• hpx::resiliency::experimental::async_replay_validate: This version of replay adds an argument to async replay which receives a user-provided validation function to test the result of the task against. If the task’s output is validated, the result is returned. If the output fails the check or an exception is thrown, the task is replayed until no errors are encountered or the number of specified retries has been exceeded.

• hpx::resiliency::experimental::async_replicate: This is the most basic implementation of the task replication. The API returns the first result that runs without detecting any errors.

• hpx::resiliency::experimental::async_replicate_validate: This API additionally takes a validation function which evaluates the return values produced by the threads. The first task to compute a valid result is returned.

• hpx::resiliency::experimental::async_replicate_vote: This API adds a vote function to the basic replicate function. Many hardware or software failures are silent errors which do not interrupt program flow. In order to detect errors of this kind, it is necessary to run the task several times and compare the values returned by every version of the task. In order to determine which return value is “correct”, the API allows the user to provide a custom consensus function to properly form a consensus. This voting function then returns the “correct”” answer.

• hpx::resiliency::experimental::async_replicate_vote_validate: This combines the features of the previously discussed replicate set. Replicate vote validate allows a user to provide a validation function to filter results. Additionally, as described in replicate vote, the user can provide a “voting function” which returns the consensus formed by the voting logic.

• hpx::resiliency::experimental::dataflow_replay: This version of dataflow replay will catch user-defined exceptions and automatically reschedules the task N times before throwing an hpx::resiliency::experimental::abort_replay_exception if no task is able to complete execution without an exception. Any arguments for the executed task that are futures will cause the task invocation to be delayed until all of those futures have become ready.

• hpx::resiliency::experimental::dataflow_replay_validate : This version of replay adds an argument to dataflow replay which receives a user-provided validation function to test the result of the task against. If the task’s output is validated, the result is returned. If the output fails the check or an exception is thrown, the task is replayed until no errors are encountered or the number of specified retries have been exceeded. Any arguments for the executed task that are futures will cause the task invocation to be delayed until all of those futures have become ready.

• hpx::resiliency::experimental::dataflow_replicate: This is the most basic implementation of the task replication. The API returns the first result that runs without detecting any errors. Any arguments for the executed task that are futures will cause the task invocation to be delayed until all of those futures have become ready.

• hpx::resiliency::experimental::dataflow_replicate_validate: This API additionally takes a validation function which evaluates the return values produced by the threads. The first task to compute a valid result is returned. Any arguments for the executed task that are futures will cause the task invocation to be delayed until all of those futures have become ready.

• hpx::resiliency::experimental::dataflow_replicate_vote: This API adds a vote function to the basic replicate function. Many hardware or software failures are silent errors which do not interrupt program flow. In order to detect errors of this kind, it is necessary to run the task several times and compare the values returned by every version of the task. In order to determine which return value is “correct”, the API allows the user to provide a custom consensus function to properly form a consensus. This voting function then returns the “correct” answer. Any arguments for the executed task that are futures will cause the task invocation to be delayed until all of those futures have become ready.

• hpx::resiliency::experimental::dataflow_replicate_vote_validate: This combines the features of the previously discussed replicate set. Replicate vote validate allows a user to provide a validation function to filter results. Additionally, as described in replicate vote, the user can provide a “voting function” which returns the consensus formed by the voting logic. Any arguments for the executed task that are futures will cause the task invocation to be delayed until all of those futures have become ready.

See the API reference of the module for more details.