Method and system for dynamic node allocation in data flow architecture

Abstract

The invention discloses a dynamic node allocation method in the data flow system structure. The method maps the nodes in the data flow graph to the physical execution unit one by one. The mapped nodes are called dead nodes, and a node is set in the data flow graph. Management node, there is a data transmission path between the management node and each node; an information table is set at the management node, and live nodes are counted in the information table; a congestion threshold is set for each dead node; whether the dead node is congested, If the dead node detects that the upstream node has completed execution but its own instructions are still being executed, and the execution time exceeds its corresponding congestion threshold, it is determined that the dead node has congestion; when a dead node A is congested, the dead node A will report to the management The node sends a request message; after receiving the request message, the management node maps the request message to a live node according to the actual situation; when the lease time of the live node B reaches the lease period, the management node sends a recycling request to the live node.

Description

technical field [0001] The present invention relates to the technical field of data flow architecture, in particular to a method and system for dynamic node assignment in data flow architecture. Background technique [0002] With the advent of the era of big data computing, the number and scale of data sets have higher and higher requirements on the processing power of processors. The data flow architecture has the characteristics of a large number of instruction-level parallelism and low power consumption of memory access, so it has been widely concerned by the society. In the data flow graph architecture, the program code can be transformed into a data flow graph. A data flow graph is a directed graph in which each node represents a piece of code in a program, and directed lines represent the flow of data between nodes. In the traditional von Neumann structure, the data transfer between instructions is interactive through the memory. After the previous instruction stores...

AI Technical Summary

Problems solved by technology

The problem with this method is that although it can solve the problem of slow execution of some key nodes, it requires a lot of experimental testing and experience to determine which nodes are key nodes, which will increase the preprocessing time

In addition, even for the data flow graph of the same application, the node fragments that cause data flow graph congestion are different at different periods of program execution

Therefore, even if the same data flow graph is statically optimized, it is still difficult to achieve a good acceleration effect during the entire period of program execution.

[0004] Based on the above problems, the preprocessing time required by the static optimization method is long, and for the same application, it is difficult to improve the efficiency of the whole time period by using the static optimization method, resulting in waste of computing resources at different times

Images