Performance Prediction for Lock-based Programs
Dongwen Zhang, Tongtong Wang, Yang Zhang · 2023
Using different locking mechanisms affects parallel programs' performance differently, and its impact on program performance is difficult to assess, which hinders researchers from rationally utilizing locking mechanisms. Moreover, there are few studies on the performance prediction of different locking mechanisms. To address this issue, this paper proposes a combination of deep feedforward neural network (FNN) and Random Forest (RF) method LockPerf to predict the performance of parallel programs with different locking mechanisms. The performance predicted in this paper is the execution time of the program. In this paper, extracting the static characteristics of the program first, then sets the variables such as the number of threads, lock type, and read/write ratio by switch statement, and finally runs the program to collect multiple samples to construct a configurable data set. A total of 9 projects are employed to evaluate the effectiveness of LockPerf Experimental results show that the average of the mean relative errors is 5.47, and the mean of the standard errors of the 95% confidence intervals is 0.13. The experiments show that LockPerf effectively predicts the performance of parallel programs with different locking mechanisms.