LLVM-based automatic performance prediction method for scientific calculation program

Abstract

The invention discloses an LLVM-based automatic performance prediction method for a scientific calculation program, and belongs to the technical field of program performance prediction. The automatic performance prediction method aims at achieving automatic analysis of the scientific calculation program and improving the accuracy of static analysis, and meanwhile program prediction time can be given finally. The automatic performance prediction method is characterized by including the steps of converting the predicted source program into an intermediate bitcode in an LLVM; analyzing the intermediate bitcode to obtain a calling time number, a circulating time number and a static branch probability for identifying a communication command MPI; carrying out mixed pile penetrating on the intermediate bitcode; carrying out code deletion on the intermediate bitcode obtained after mixed pile penetrating to carry out optimizing; running the optimized intermediate bitcode to obtain llvmprof.out files; analyzing the llvmprof.out files, and predicting execution time in cooperation with command time. The automatic performance prediction method is suitable for predicting the performance of the scientific calculation program.

Description

technical field [0001] The invention discloses a scientific calculation program automatic performance prediction method, which belongs to the technical field of program performance prediction. Background technique [0002] High-performance computing is an important branch of computer science, and performance is a key characteristic of high-performance computing. Program execution time is the performance feature that users are most concerned about. The reason why users choose to use hundreds of processors to process programs is that they need to obtain program running results within a limited time. Therefore, predicting the execution time of parallel programs on a certain platform has been more and more researched, and this technique is called program profiling. [0003] Performance evaluation methods can be divided into dynamic analysis and static analysis. [0004] The so-called dynamic analysis method is to predict large-scale through small-scale, that is, to measure mul...

AI Technical Summary

Problems solved by technology

If it is to measure the 1024 degree of parallelism, and the selected sampling scale range is only a curve fitted by a small range such as [0-128], it is difficult to guarantee that it is reasonable under the 1024 scale

Even if the fitting function can overlap well in the previous range, there is no guarantee that it can fit well on a large scale

This limits the scope of application of the forecast

Moreover, the parallelism of some programs is required and cannot be changed continuously. In this way, the sampled data points cannot be evenly distributed.

[0008] The results are not reasonable enough: because the whole fitting process is analyzed purely mathematically, many internal dependencies of the program are ignored, so the need for program features is also ignored

[0011] Insufficient rationality: the branch probability is only the type of prior probability and branch instruction applied, and it cannot actually represent the target program

However, this does not meet our requirements

Images