Table 4 : Resources on HPC training, resources, community
In Table 2, we mentioned the development of parallel algorithms, which
led to a transformation in multiphysics simulations: some examples
include parallel matrix operations and linear algebra
(https://cvw.cac.cornell.edu/APC/); parallel implementation of the
N-body problem with short-range interactions
(https://cvw.cac.cornell.edu/APC/); long-range interactions and
the parallel particle mesh Ewald sum [49]; parallel Monte Carlo
[50]; linear-scaling methods such as multipole expansion [49];
linear-scaling density functional theory [51]; parallel graph
algorithms (https://cvw.cac.cornell.edu/APC/). As a specific
example, we note that the N-body problem is an essential ingredient in
MD. A common goal in MD of large systems is to perform sufficient
sampling of the combinatorially large number of conformations available
to even the simplest of biomolecules [52, 53]. In this respect, a
potential disadvantage of molecular dynamics calculations is that there
is an inherent limitation upon the maximum time step used for the
simulation (≤ 2 fs). Solvated systems of biomolecules typically consist
of 105-106 atoms. For such system
sizes, with current hardware and software, simulation times extending
into the tens of microseconds regime is an exceedingly labor-intensive
and challenging endeavor that requires a combination of algorithmic
enhancements as well as the utilization of high-performance computing
hardware infrastructure. For example, cutoff distances reduce the number
of interactions to be computed without loss of accuracy for short-range
interactions but not for long-range (electrostatic) interactions;
long-range corrections such as the particle mesh Ewald algorithm
[54] along with periodic boundary conditions are typically
implemented for maintaining accuracy. Parallelization techniques enable
the execution of the simulations on supercomputing resources such as
4096 processors of a networked Linux cluster. Although a cluster of this
size is a big investment, its accessibility is feasible through the
extreme science and engineering discovery environment (XSEDE) for
academic researchers. XSEDE resources
(www.xsede.org) currently include petaflop
of computing capability, and other US national laboratories such as the
Oakridge are moving towards exascale computing
(https://www.exascaleproject.org) [55]. Another approach,
capitalizing on advances in hardware architecture, is creating custom
hardware for MD simulations, and offers one-two orders of magnitude
enhancement in performance; examples include MDGRAPE-3 [56, 57] and
ANTON [58, 59]. Graphical processing unit (GPU) accelerated
computation has recently come into the forefront to enable massive speed
enhancements for easily parallelizable tasks with early data indicating
that GPU accelerated computing may allow for the power of a
supercomputing cluster in a desktop, see e.g., [60, 61].