SC '21 | Hybrid, Scalable, Trace-Driven Performance Modeling of GPGPUs

[mental2008]

Presented in SC '21. [ Paper ]

Authors: Yehia Arafa, Abdel-Hameed Badawy, Ammar ElWazir, Atanu Barai, Ali Eker, Gopinath Chennupati, Nandakishore Santhi, Stephan Eidenbenz New Mexico State University, State University of New York at Binghamton, Amazon Alexa, Los Alamos National Laboratory

Code

1. PPT-GPU: https://github.com/NMSU-PEARL/PPT-GPU
2. GPGPUs-Workloads: https://github.com/NMSU-PEARL/GPGPUs-Workloads

[mental2008]

The authors present a scalable performance prediction toolkit (PPT-GPU) for NVIDIA GPUs. Its modeling methodology uses a hybrid combination of analytical and event-driven modeling approaches. It can be easily extended to model the new features in contemporary GPUs (interesting, how to achieve this?). It has been verified against NVIDIA profiling tools (nv-nsight).

The overview of PPT-GPU is as follows:

• Memory Model: predict memory metrics such as memory requests, transactions, hit rates, and divergence.
• Compute Model: predict computational metrics such as cycles, instructions executed, and occupancy.