Synchronous deployment system and method for multi-stream parallelism

Abstract

The invention relates to a synchronous deployment system and method for multi-stream parallel parallelism The system comprises: an initial calculation graph obtaining component used for obtaining an initial calculation graph comprising a plurality of task flows belonging to a calculation task, each task flow comprises a plurality of operation logic nodes which are executed in sequence, and each operation logic node comprises a task flow mark which the operation logic node belongs to; a node relation analysis component used for traversing all tensors to be targeted in the computational graph, and judging whether task flows to which a first node targeted to a first tensor and a second node targeted to the first tensor belong for the first time are the same or not; and a synchronization node deployment component used for deploying a first synchronization dependence node marking the first task flow before the second node when the node relationship analysis component determines that the first node belongs to the first task flow and the second node belongs to the second task flow different from the first task flow, marking the second node with the first synchronization dependence node, and determining that the operation of the first node is prior to the operation of the second node through the first synchronization dependent node.

Description

technical field [0001] The present disclosure relates to a data processing technique. More specifically, the present disclosure relates to a multi-stream parallel synchronous deployment system and method thereof. Background technique [0002] Now that deep learning is popular, in order to increase the speed of data processing, it is usually necessary to divide a task into multiple slice tasks to form multiple task flows, so that the task flows can be processed in parallel, thereby saving data processing time or speeding up tasks. Processing efficiency, this data processing method is multi-stream parallel processing. [0003] However, in the multi-stream parallel data processing method, there are often parameter exchanges between different streams and data dependencies between different streams. Therefore, if parameters are not synchronized in parallel, errors will occur when the results of each task flow are merged. This asynchrony is often caused by different task volume...

AI Technical Summary

Problems solved by technology

Due to two Kernels, S1-n1 and S2-n2, in two different task flows S1 and S2, the execution of S2-n2 may be earlier than S1-n1, in this case, S2-n2 reads data may be wrong

On the other hand, for Kernels that are logically undependent on multiple parallel task flows, it is possible that the tensor life cycle between different Kernels is uncontrollable, resulting in memory data errors

For example, three Kernels, such as S3-n3, S2-n2, and S1-n2, are in different task flows S3, S2, and S1 respectively. S3-n3 and S2-n2 need to read the tensor T2 in the memory M, In order to achieve synchronization, the conventional method will unconditionally instruct S3-n3 and S2-n2 to wait for S1-n2 or to synchronize with S1-n2. These synchronizations are adjusted manually, which is very labor-intensive

Moreover, when there are multiple parallel task flows that need to read the tensor T2 in the memory M, there will be multiple waiting or synchronization processes, resulting in redundant synchronization operations, resulting in complicated data processing

Images