Parallel and Distributed Processing for<tex>$hhbox-p$</tex>Adaptive Finite-Element Analysis: A Comparison of Simulated and Empirical Studies

A New Strategy for Reducing Communication Latency in Parallel 3-D Finite Element Tetrahedral Mesh Refinement

“…7(a) describes the regular parallelization speedup without pipelined communication: using 3-8 PEs can yield a speedup of 1.45 2.15 times faster than one PE. For the initial iterations, 5-6 PEs yields a better speedup than 7-8 PEs, because the message size is too small relative to the parallel overhead [2]. Fig.…”

“…7(b), 6PEs performed better than 5, 7, or 8 PEs. The reason is increasing the number of breakpoints can reduce the probability of a collision between any of two slave PEs in the data transmission, but when the number of work segments is increased, the pipelining cost and communication overhead are also increased accordingly [2]. This is a beneficial tradeoff required to avoid the block points in the system.…”

Efficient load balancing for parallel adaptive finite-element electromagnetics with vector tetrahedra

“…A comparison of the solution times of the various approaches considered above with that of a basic single machine serial solution is given in Table V. The results are based on a symmetric TABLE V SPEED-UP RELATIVE TO SINGLE MACHINE SERIAL SOLUTION TIME multiprocessor with four identical processors (Sun v880) [3]. The load balancing achieved by the new approach results in improved solution times relative to the other methods (26% and 53% relative to the Uniform/Adaptive and Uniform/Nonadaptive approaches, respectively).…”

“…However, the electromagnetic simulation of some complex problems can still be intractable because of the very large number of degrees of freedom (DOF) necessary for sufficiently accurate solutions. One approach for overcoming this barrier is to combine AFEMs with high-performance computing (HPC) methods such as parallel and distributed simulations [2], [3].…”

“…Hence, some subdomains will require more refinement than others as the discretization is adaptively refined. Consequently, an initial discretization that is evenly distributed among processors in terms of workload can lead to a severe workload imbalance as the adaption progresses toward more nonuniform discretizations [2], [3]. This makes the task of balancing computational workload among available processors, in a cost-effective manner, nontrivial.…”