R&D on FRIB on-line fast lattice optimizer using cloud technique

Project Details

Project Lead
Zhengqi He 
Project Manager
Zhengqi He 
Michigan State University, FRIB  
Engineering Science and Engineering Physics (107) 
14.12 Engineering Physics 


My aim is to develop a system which is capable of controlling all beam line elements while communicating with diagnostic system, machine protection system and any other related systems. The system must be able to transport all kinds of ion beams to target with designed beam power and beam energy with minimum beam loss; stable and error robust; fast transition between different ion species; adaptive to varies error and fault conditions. The most difficult part for the system lies in On-line particle tracking based optimization (with detected error and fault included) and machine error and machine fault prediction (usually based on particle tracking results). These two applications form a close loop and would be important for the system to become error and fault robust. However, particle tracking is time consuming, the typical time for using optimizer and particle tracking program to give a feasible lattice is 1 week. The typical grace period of on-line large scaling computing is within hours, so the calculation speed is still far from enough. The only solution to this problem is high performance computation. I’ve been considering an MPI-Cloud or (GPU-Cloud) based mixed high performance computation infrastructure, this infrastructure would not only suitable for running particle tracking program, beneficial to FRIB project, but also suitable for all kinds of large scale Monte-Carlo simulation or case-by-case searching optimization.

Intellectual Merit

(1) Extend Cloud Computing technique to accelerator engineering. (2) Provide new solution to accelerator on-line tuning strategy and design strategy. (3) Explore the most effective way of using cloud computing technique on Monte-Carlo like problem.

Broader Impacts

The research tries to transplant cloud computing technology to accelerator physics realm, providing new possibility and strategy on on-line tuning and accelerator lattice design problem.

Scale of Use

5 nodes at a time on developing phase, quite frequently. And 30 nodes at a time for performance check, quite rare. The simulation code can take 3 minutes per time and 30 times for developing phase. 2000 times are needed for optimization performance check.