Using colored petri nets for GPGPU performance modeling

Abstract

Performance analysis and modeling of applications running on GPUs is still a challenge for most designers and developers. State-of-the-art solutions are dominated by two classic approaches: statistical models that require a lot of training and profiling on existing hardware, and analytical models that require in-depth knowledge of the hardware platform and significant calibration. Both these classes separate the application from the hardware and attempt a high-level combination of the two models for performance prediction. In this work, we propose an orthogonal approach, based on high-level simulation. Specifically, we use Colored Petri Nets (CPN) to model both the hardware and the application. Using this model, the execution of the application is a simulation of the CPN model using warps as tokens. Our prototype implementation of this modeling approach demonstrates promising results on a few case studies on two different GPU architectures: both reasonably accurate predictions and detailed execution information are obtained. We conclude that CPN-based GPU performance modeling is an elegant solution for systematic performance prediction, and we focus further on optimizing the models to improve the execution time of the symbolic simulation.

Modeling of GPGPU architectures for performance analysis of CUDA programs

“…The average error measurement is 6%, comparable to other works in literature, e.g., in [20], the average error measurement is 6% on the NVIDIA FERMI GTX 480 and 8% on the NVIDIA KEPLER K20m. If we measure the error from 32x32 to 256x256, obtain an average error of 0.0275.…”

“…Hence, if one is interested only in estimating the number of clock cycles required to run an application, it is sufficient to properly set this parameter. This approach is consistent with other works on GPU stochastic modeling, discussed in Section 2.2 [20], [21]. For a resource-oriented analysis, it is also possible to estimate metrics for each hardware resource represented in our model.…”

“…However, their focus is on modeling the whole map-reduce process, and they do not detail the execution of the program's instructions on the GPU, nor they do consider the GPU's structure and the interactions between its components. The authors of [20] use Colored Petri Nets to evaluate the performances of GPGPUs, using colors to represent different instruction types. They model the behavior of two different kernels on two separate GPU's architectures.…”

“…Thanks to the use of the NVIDIA Nsight profiler, it is also possible to know how many times a specific instruction was executed. This is extremely useful to provide a realistic and faithful representation of the program's execution, to handle situations such as branch divergency, and it can also be obtained from static code analysis as shown in [20], [21]. Each SASS instruction is associated to an identifier representing the kind of operation, e.g., ALU operation or Load/Store.…”

High-level hardware feature extraction for GPU performance prediction of stencils

“…COM-PASS [18] introduces a language for creating analytical performance models that analyze the amount of loating point and memory operations based on static code features. Coloured petri nets [20] were proposed for GPGPU performance modelling. Another approach [3] builds an analytical performance model to determine the lower bound on execution time.…”