Biomedical Engineering Reference
In-Depth Information
starting points for the replica simulations and data are collected after 100 ps REMD
simulations. Each replica is run for 3.0 ns for data collection, with replica exchanges
attempted every 0.4-2 ps [15, 26, 35, 36].
17.3.2 Optimal Temperature Sequences in REMD
The optimal temperature distributions in the replica exchange method can be obtained
by running a few trial replicas with short MD simulations. The temperature gap can
thus be determined by monitoring the acceptance ratio desired between neighbor-
ing temperatures. The rest of the temperature list can usually be interpolated, as the
optimal temperature distribution should be roughly exponential assuming the heat
capacity is relatively a constant [15]. Using the
β
-hairpin as an example, the optimal
temperature sequence is found to be 270, 274, 278,
...
, 685, and 695 K, with an accep-
tance ratio of about 30-40%. The temperature gaps between these replicas range from
4 to 10 K. Both temperature swap and configuration swap (coordinates and velocities)
have been implemented in REMD with MPI, and very high efficiency (up to 98%)
has been achieved with this embarrassingly parallel algorithm. Figure 17.1 shows the
speedup for the example
β
-hairpin with 1, 2, 4,
...
, 64, 128 processors or replicas
(each replica is run on one processor). The speedup is measured by the aggregate
MD steps (number of replicas multiplied by the MD steps in one replica) finished by
Efficiency of REMD
100
10
1
1
10
100
Number of processors
Figure 17.1
The speedup of REMD versus the number of processors used. The speedup is
measured by the aggregate MD steps (number of replicas multiplied by the MD steps in one
replica) finished by REMD within 1 h of wall-clock time divided by that finished by a single
processor and single replica run. The simulations were done on the solvated
β
-hairpin system
with replica exchanges attempted every 250 MD steps.
Search WWH ::
Custom Search