background image
c CCLRC
Section 4.3
forces). (Some communications have compulsory character and therefore higher priority to calcu-
lations that are to follow.) In such regimes, the overall DL POLY 3 efficiency falls down since
processors stay idle and no power is harnessed while communications are not finished.
It is important that the user recognises when DL POLY 3 becomes vulnerable to decreased effi-
ciency and what possible measures could be taken to avoid this. DL POLY 3 calculates and reports
the link-cell algorithms (M
x
· M
y
· M
z
) employed in the simulations immediately after execution.
M
x
(analogously for M
y
and M
z
) is the integer number of the ratio of the width of the system
domains in x-direction to the maximal short-range cutoff specified for the system:
M
x
= Nint
Max(x
ij
)/#(nodes)
x-direction
Max(cutoff)
.
(4.1)
Every domain (node) of the MD cell is loaded with (M
x
+ 2) · (M
y
+ 2) · (M
z
+ 2) link-cells
of which M
x
· M
y
· M
z
belong to that domain and the rest are a halo image of link-cells from
neighbouring domains. In this respect, the more linked cells per domain, the less halo data to keep,
the more efficient the load distribution per node and the less the communications between nodes
and therefore, the better the parallelisation.
DL POLY 3 issues a built-in warning when a link-cell algorithms has a dimension less than four (i.e.
less than four link-cells per domain in given direction). A useful rule of thumb is that parallelisation
speedup inefficiency is expected when the ratio:
R =
M
x
· M
y
· M
z
(M
x
+ 2) · (M
y
+ 2) · (M
z
+ 2) - M
x
· M
y
· M
z
is close to or drops below one. In such cases there are three strategies for improving the situation
that can be used singly or in combination. As obvious from equation (
4.1
) these are; (i) decrease
the number of nodes used in parallel, (ii) decrease the cutoff and (iii) increase system size. It is
crucial to note that increased parallelisation efficiency remains even when the link-cell algorithm is
used inefficiently. However, DL POLY 3 will issue an error message and cease execution if detects
it cannot fit a link-cell per domain as this is the minimum the DL POLY 3 link-cell algorithm can
work with - (1 · 1 · 1) corresponding to ratio R = 1/26.
Performance may also affected by the fluctuations in the inter-node communication, due to un-
avoidable communication traffic when a simulation job does not have exclusive use of all machine
resources. Such effects worsen the performance very much when the real time-per-timestep is of the
same magnitude as the average communication time (i.e. nodes spend more time communicating
rather than calculating) which is specific for the specific simulation control parameters.
4.3
A Guide to Preparing Input Files
The CONFIG file and the FIELD file can be quite large and unwieldy particularly if a polymer
or biological molecule is involved in the simulation. This section outlines the paths to follow
when trying to construct files for such systems. The description of the DL POLY 3 force field in
Chapter
2
is essential reading. The various utility routines mentioned in this section are described
in greater detail in the DL POLY 2 Reference Manual. Many of these have been incorporated into
the DL POLY 3 Graphical User Interface [
15
] and may be conveniently used from there.
4.3.1
Inorganic Materials
The utility genlat can be used to construct the CONFIG file for relatively simple lattice structures.
Input is interactive. The FIELD file for such systems are normally small and can be constructed
68