I'm working with Jonathan Wurtele and Ji Qiang at LBNL's Center for Beam Physics to simulate a next generation light sources based on x-ray free electron lasers.
My research includes:
- Designing parallel algorithms for numerical optimization.
- Profiling and performance enhancements for the IMPACT-T particle accelerator code.
- Contributing to a reduced physics/cost model of an FEL facility.
Performance Enhancements for IMPACT-T
IMPACT-T simulates particle accelerators by tracking the 3-D motion of relativistic particles through the external electric and magnetic fields of the accelerating structure. Two computational phases are most important to IMPACT's performance. The "kick" phase updates the particles'‚ position and momentum, and is embarrassingly parallel if properly load-balanced. The "solver" phase uses Particle-In-Cell (PIC) approach to compute space-charge forces. The transpose steps of IMPACT‚'s FFT-based Poisson solver require all-to-all communication and may be improved by alternative communication patterns.
I have developed two alternative communication strategies ("overlapped" and "packed") for the 3D-FFT and compared them to the original IMPACT implementation ("serial"). The new "packed" algorithm is fastest in all cases (especially at high concurrency), and is up to 40% faster than the original version.