Performance Analysis of Load Balancing Algorithms for Parallel SAMR Application

Henrik Johansson
Division of Scientific Computing
Department of Information Technology
Uppsala University


Abstract:

Structured adaptive mesh refinement (SAMR) is widely used to decrease the run time of simulations that solve partial differential equations (PDEs). Employing SAMR, the run time can often be reduced to a small fraction. SAMR is used in areas like computational fluid dynamics, numerical relativity, astrophysics, and hydrodynamics.

To efficiently use SAMR on parallel computers, the dynamic grid hierarchy must repeatedly be repartitioned and redistributed over the processors. The partitioning process must not only take the computations into account but also all other factors that affect the run time. A significant portion of the run time consists of communication, synchronization and data movement. The run time is further influenced by the performance and utilization of the network. Thus, to minimize the run time, both the current state of the application and the underlying hardware must be taken into account. This is non-trivial, since the basic conditions for how to allocate the resources change dramatically during run-time due to the dynamics inherent in both the applications and the computer system.

In this seminar, a performance analysis of SAMR load balancing algorithms is presented. A common partitioning algorithm is, on a time-step basis, compared against a large number of advanced partitioning algorithms. For the comparisions, four real-world applications are used. The data collected during this process will incorporated into the meta-partitioner, which autonomously selects, configures, and invokes the best partitioning algorithm with regard to the current application and computer state.