Graphical processor-based graph computing edge vector load balancing method and device

Abstract

The invention discloses a graph calculation edge vector load balancing method and device based on a graphics processor. The method of the invention includes regularly finding out the maximum and minimum amount of data among the edge data that all SM cores need to process and calculating the difference, And the difference exceeds the preset threshold and the states of the two SM cores are respectively set to the state of waiting for the donated load and receiving the donated load, the control processor of the graphics processor sends a load balancing request to the SM core waiting for the donated load, and upon receiving The start index and end index of the donation data returned by the SM core are sent to the SM core receiving the load donation, and the load receiver records it as the side data to be processed. The present invention can efficiently realize graph-oriented computing edge vector load balance between stream processor cores of graphics processors, has the advantages of simple hardware structure and transparent programming for programmers, and can complete all operations without modifying the original program. load balancing tasks.

Description

technical field [0001] The invention relates to a load balancing technology of a graphic processor, in particular to a method and a device for balancing a graphic processor-based graph calculation side vector load. Background technique [0002] Modern microprocessor architectures continue to increase the parallel processing capabilities of microprocessors in order to achieve higher throughput rates while maintaining good energy efficiency. Graphics processing unit (GPU) is currently the computing platform with the strongest parallel processing capability, which can provide huge computing power while using low energy consumption. GPU programming languages ​​allow programmers to execute applications in parallel using hundreds or thousands of threads, such as CUDA and OpenCL. Each thread will execute the same piece of kernel code according to the single instruction multiple thread (SIMT) programming model. This SIMT architecture provides huge parallel computing capability and...

AI Technical Summary

Problems solved by technology

However, the software load balancing strategies that have been proposed have some common disadvantages: 1) These strategies need to modify the source code and add additional complex code to achieve side vector load balancing, which will make it unusable for users who are not familiar with GPU programming, The use threshold has been raised; 2) Due to the use of complex edge vector load balancing strategies, a large amount of time is required for preprocessing calculations. For graphs with uniform distribution of node degrees, this part of the time consumption does not obtain higher benefits and The performance will be much lower than the simple load balancing strategy; (3) The extra computing overhead brought by the software edge vector load balancing strategy is very large, resulting in a significant increase in the number of executed instructions

Images