Collaborative Research: SHF: Medium: Enabling GPU Performance Simulation for Large-Scale Workloads with Lightweight Simulation Methods

点击翻译按钮获取中文摘要

Designing the architecture of new computer chips typically relies on detailed simulations to avoid expensive manufacturing processes. However, the speed of computer architecture simulations has not kept up with the rapid advancements in computing technology, particularly for systems that execute applications with large computations, large memory requirements, and large communication needs. This project introduces innovative lightweight simulation techniques that focus on efficiency by selectively simulating certain aspects of chip design or using higher levels of abstraction, drastically speeding up the simulation process. This project will enable the research community with techniques that support quicker development of new computing technologies. The research outcome will make the field of computer architecture more accessible to researchers with fewer resources. Moreover, the simulation techniques derived from this project will be integrated into the computer architecture curricula, helping students, especially under-resourced students, better understand concepts related to large-scale computing. Traditional computer architecture simulators recreate cycle-by-cycle details of the hardware execution, hindering fast simulation. To improve performance, this project introduces a novel suite of simulation tools designed to support the design and optimization of next-generation, large-scale computing systems. The approach encompasses three complementary strategies: behavior modeling, sampled simulations, and data-driven simulation. Behavior modeling abstracts hardware components to focus on essential performance metrics, enabling faster simulations without significant loss of accuracy. Sampled simulations leverage the repetitive nature of applications (with a special focus on GPU applications) to predict performance by simulating only critical segments of the workload. Data-driven simulations take advantage of statistical and performance modeling techniques to further advance simulation capabilities. These strategies will be unified under the Akita simulator framework, facilitating interoperability and ease of use across different simulation schemes.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.