Relativistic Particle Transport Systems
The system with large number of particles can be described by the phase space distribution, a snapshot of where particles are and how they move. The system does not have to be in thermal equilibrium. In other words, the system evolution can start with any initial phase-space distribution. The late time distribution can be obtained by solving the Boltzmann equation. When the system is in equilibrium and is able to maintain equilibrium, the Boltzmann equation can lead to macroscopic hydrodynamical equations. In addition, the Boltzmann equation can be used for the study of equilibration from non-equilibrium initial conditions and for the study of systems far from equilibrium. This enables research beyond equilibrium and near-equilibrium situations.
The Boltzmann equation will be solved by the test particle method. The phase-space will be sampled by point particles (delta functions in phase-space). These point particles will move in straight lines between collisions. When two particles come to their closest distance, the closest distance is compared with the interaction range. If the closest distance is smaller than the interaction range, there will be a collision. Otherwise there is no collision between these two particles.
A partonic system is dominated by light (massless) partons. These massless
partons are ultrarelativistic, i.e., they move at the speed of light. Different
from classical billiard balls, space and time are related relativistically. The
closest distance should be calculated in the two colliding particle
center-of-mass system in which the total momentum of the two colliding
particles is zero. The collision space-time point is taken to be the midpoint
at closest approach time in the two-particle center-of-mass system.
Speculative discrete event simulation has been considered for this system. Particle collisions are calculated aggressively and some correctly guessed ones would save the simulation execution whereas those over-optimistic ones will be eliminated later. Multithreaded simulation program has been developed in multicore and SMP clusters. Both computational and communication operations are overlapped. Checkpointing strategy has been applied for fault resilience. Specially developed 3-D localized load balancing scheme is utilized to reduce rescheduling overhead.
Heterogeneous processing architecture with GPUs (Graphics Processing Units) has been considered to take advantage of the massive computing cores in GPUs. However, collision dependency exhibits as one of major obstacles to employ this SPMD (Single Program Multiple Data Streams) programming paradigm. Divergence and memory-space-selection strategy are the major issues. Further investigation is undergoing.
Dynamic thread scheduling in GPU clusters will be the ultimate challenge to orchestrate data and computations among multi-level heterogeneous processing units. Workloads can be adjusted automatically to utilize system resources thoroughly. The execution will be sped up whereas system throughput will be improved. Similar strategy can be applied on other scientific computing applications.
Investigator: Jiang (ASU)