Parallel CBMC

Another important application of the CBMC algorithm is the calculation of the chemical potential using Widom s test particle method [29,32,96], For the parallel CBMC algorithm described here it is straightforward to show that the excess chemical potential can be calculated using... 

To test this parallel algorithm, we have studied a system of a single n-hexane molecule in the zeolite Silicalite. Details about the model we used can be found in appendix A. Our simulation code was written in FORTRAN77 cind peirallelized using MPI [98]. The following machines were used to test our parallel CBMC algorithm ... 

In this appendix, we will prove that our parallel CBMC algorithm obeys detailed balance. We start with our super-detailed balance [29] equation ... 

A completely different application of the CBMC ideas is used by Esselink et al. to develop an algorithm to perform Monte Carlo simulations on parallel computers... 

There are a number of reasons why the conventional CBMC algorithm cannot be parallelized efficiently ... 

When the calculation of 6ui and 6u2 is computationally expensive cmd not easy to calculate in parallel, IN-DC-CBMC will fcdl because too many correction terms have to be calculated and OUT-DC-CBMC will fail because the calculation of 6u2 cannot be parallelized. In appendix C, we present an algorithm that may solve this problem. 

In summary, we have briefly reviewed some of the interesting aspects of CBMC. It was found that we can efficiently use parallel computers to perform CBMC simulations and that special care has to be taken for the simulation of branched molecules. 

In summary, we have extended the recoil growth scheme for systems with continuous potentials. We find that in a NVT simulation RG is much more efficient thcin CBMC for long chains and high densities. However, in appendix B we have shown that RG is less suitable for parallelization using the parallel algorithm of section 2.3. We found that the standard Metropolis acceptance/rejection rule is a reasonable stochastic rule when a Lennard-Jones potential is used. 

Jones potential with a = e = (3 = 1, this corresponds to a pair separation of = 0.64. It turns out that the distribution in CPU time is much wider for RG than for CBMC. The fact that in RG many configurations can be thrown away before the complete chain is constructed is one of the main reasons why RG is more efficient than CBMC. However, it also implies that multiple chain algorithms cannot be parallelized efficiently, which makes RG less suitable for parallelization. 

Vlugt, T.J.H., and Smit, B. (2000). Advanced CBMC techniques . In Proceedings of the Workshop Molecular Dynamics on Parallel Computers , 8-10 February 1999, Julich, Germany, Editors R. Esser, P. Grassberger, J. Grotendorst, M. Lewerenz, World Scientific, 2000. 

---