SPEAR

Systems for Performance, Energy, and Resiliency

The SPEAR (Systems for Performance, Energy, and Resilience) team envisions a future in which high-performance computing (HPC) and artificial intelligence (AI) evolve together to create intelligent, adaptive, and sustainable computing systems that accelerate scientific discovery and technological innovation.

Our research spans the foundations of next-generation computing systems, including cluster and resource management, digital twins, performance modeling and simulation, networking, power and energy efficiency, and resilience. Building on these foundations, we pursue the convergence of HPC and AI through two complementary and mutually reinforcing directions. Through HPC4AI, we design scalable, high-performance, and energy-efficient systems that enable increasingly capable AI models and applications. Through AI4HPC, we harness AI to transform how complex computing systems are designed, operated, optimized, and made resilient.

Members

• Zhiling Lan (Professor)

• Mike Papka (Professor)

• Xin Wang (Postdoc at UIC)

• Yuping Fan (Postdoc at ANL)

• Melanie Cornelius (PhD)
• Matthew Dearing (PhD)
• Zhong Zheng (PhD)
• Greg Cross (PhD)
• Yash Kurkure (PhD)
• Chris Grams (PhD)
• Amy Byrnes (PhD)
• Yihe (Jordan) Zhang (PhD)
• Yiheng Tao (PhD)
• Maisy Dunlavy (PhD)
• Kanglin Xu (PhD)
• Can Bagirgan (PhD)
• Aldo Cabrera (PhD)
• Akshar Patel (PhD)
• Niccolo Brembilla (PhD)
• Giacomo Brunetta (PhD)
• Vincent Chirio (PhD)
• Jennifer Coburn (PhD)
• Edoardo D’Alessio (PhD, co-advised with Prof Xiaoguang Wang)

• Hao Liu (PhD)

• Kim Meagher (MS)
• Vittorio Palladino (MS)

Collaborators

We have a close partnership with several research teams at ANL, including the ALCF Operations team and the performance team led by Valerie Taylor.

Software

We actively develop and maintain open-source software on SPEAR GitHub. Several representative tools are listed below:

• CQSim: an event-driven scheduling simulator designed for rapid what-if exploration of scheduling scenarios.
• Q-adaptive: a multi-agent reinforcement learning driven routing design for Dragonfly networks.
• MFNetSim: a hybrid network simulation framework for joint MPI and I/O modeling on Dragonfly systems.
• MAGUS: a system-level library for adaptive uncore scaling that minimizes power waste on heterogeneous CPU-GPU systems.
• SMART: a hybrid GNN+LLM surrogate model for predicting application runtime on Dragonfly systems.
• DNPC: a user-level dynamic power capping library for parallel applications.

SPEAR in Action