| 1 | #!/usr/bin/env python
|
|---|
| 2 |
|
|---|
| 3 | #import smmp
|
|---|
| 4 | import sys, copy, random, math, numpy as np
|
|---|
| 5 |
|
|---|
| 6 | from mpi4py import MPI
|
|---|
| 7 | #from algorithms import CanonicalMonteCarlo
|
|---|
| 8 |
|
|---|
| 9 | class ParallelTempering(object):
|
|---|
| 10 | """Parallel tempering is a generalized Monte Carlo algorithm. A set of
|
|---|
| 11 | replica is simulated at different temperatures. Each replica performs a
|
|---|
| 12 | canoncial Monte Carlo simulation. After a given number of sweeps, an exchange
|
|---|
| 13 | is attempted between neighboring replicas.
|
|---|
| 14 | """
|
|---|
| 15 |
|
|---|
| 16 | def __init__(self, T, myUniverse):
|
|---|
| 17 | """Initialize ParallelTempering with a set of tempertures T.
|
|---|
| 18 | """
|
|---|
| 19 | self.kB = 1.98773e-3
|
|---|
| 20 |
|
|---|
| 21 | self._T = copy.deepcopy(T)
|
|---|
| 22 | self._numberOfReplica = len(T)
|
|---|
| 23 |
|
|---|
| 24 | # Calculate the number of processors available per replica
|
|---|
| 25 | numberOfProcessors = MPI.COMM_WORLD.Get_size()
|
|---|
| 26 | numberOfProcessorsPerReplica = numberOfProcessors // self._numberOfReplica
|
|---|
| 27 | numberOfProcessors = numberOfProcessorsPerReplica * self._numberOfReplica
|
|---|
| 28 | self._myReplica = MPI.COMM_WORLD.Get_rank() // numberOfProcessorsPerReplica
|
|---|
| 29 |
|
|---|
| 30 | # Create the replica internal communicators for the energy function
|
|---|
| 31 | groupWorld = MPI.COMM_WORLD.Get_group()
|
|---|
| 32 | group = [MPI.Group.Range_incl(groupWorld, [[i * numberOfProcessorsPerReplica, (i + 1) * numberOfProcessorsPerReplica - 1, 1], ]) for i in xrange(self._numberOfReplica)]
|
|---|
| 33 | comm = [MPI.COMM_WORLD.Create(g) for g in group]
|
|---|
| 34 |
|
|---|
| 35 | # Create the communicator needed to exchange information about replica
|
|---|
| 36 | ptGroup = MPI.Group.Incl(groupWorld, range(0, numberOfProcessors, numberOfProcessorsPerReplica))
|
|---|
| 37 | self._ptComm = MPI.COMM_WORLD.Create(ptGroup)
|
|---|
| 38 |
|
|---|
| 39 | self._localMaster = self._ptComm != MPI.COMM_NULL
|
|---|
| 40 | if self._localMaster:
|
|---|
| 41 | self._myPTRank = self._ptComm.Get_rank()
|
|---|
| 42 |
|
|---|
| 43 | self._oddCycle = False
|
|---|
| 44 |
|
|---|
| 45 |
|
|---|
| 46 | def exchange(self):
|
|---|
| 47 | """Collect temperatures and energies from all replicas, determine neighbor and do replica exchange move.
|
|---|
| 48 |
|
|---|
| 49 | The exchange step is the heart of parallel tempering. The probability of exchange between two
|
|---|
| 50 | temperatures i and j is given by P_{ij} = \Delta \beta \Delta E, where \beta = 1/k_B T.
|
|---|
| 51 | """
|
|---|
| 52 | if self._localMaster:
|
|---|
| 53 | # energies = ptComm.allgather(self._myUniverse.energy())
|
|---|
| 54 | myE = [250, 255, 75, 125, 120, 185][self._myReplica]
|
|---|
| 55 | myT = [420, 370, 250, 275, 300, 330][self._myReplica]
|
|---|
| 56 | energies = self._ptComm.allgather(myE)
|
|---|
| 57 | # temperatures = ptComm.allgather(self._myUniverse.temperature())
|
|---|
| 58 | temperatures = self._ptComm.allgather(myT)
|
|---|
| 59 |
|
|---|
| 60 | temperatureRank = zip(temperatures, np.arange(self._numberOfReplica))
|
|---|
| 61 | temperatureRank.sort()
|
|---|
| 62 |
|
|---|
| 63 | if (self._myPTRank % 2):
|
|---|
| 64 | if self._oddCycle:
|
|---|
| 65 | rank0 = self._myPTRank
|
|---|
| 66 | else:
|
|---|
| 67 | rank0 = self._myPTRank - 1
|
|---|
| 68 |
|
|---|
| 69 | rank1 = (rank0 + 1) % self._ptComm.Get_size()
|
|---|
| 70 |
|
|---|
| 71 | self._oddCycle = not self._oddCycle
|
|---|
| 72 |
|
|---|
| 73 | E0 = energies[temperatureRank[rank0][1]]
|
|---|
| 74 | E1 = energies[temperatureRank[rank1][1]]
|
|---|
| 75 | T0 = self._T[rank0]
|
|---|
| 76 | T1 = self._T[rank1]
|
|---|
| 77 | beta0 = 1.0 / (T0 * self.kB)
|
|---|
| 78 | beta1 = 1.0 / (T1 * self.kB)
|
|---|
| 79 |
|
|---|
| 80 | P = math.exp((beta0 - beta1) * (E0 - E1))
|
|---|
| 81 | print self._myPTRank, P
|
|---|
| 82 |
|
|---|
| 83 | if P >= random.random(): # accept exchange
|
|---|
| 84 | rank = self._ptComm.allgather(np.array([temperatureRank[rank1][1], temperatureRank[rank0][1]]))
|
|---|
| 85 | else:
|
|---|
| 86 | rank = self._ptComm.allgather(np.array([temperatureRank[rank0][1], temperatureRank[rank1][1]]))
|
|---|
| 87 | else:
|
|---|
| 88 | random.random() # Do one random number to stay in sync with the other processes for the same replica.
|
|---|
| 89 | rank = self._ptComm.allgather(np.array([-1, -1]))
|
|---|
| 90 |
|
|---|
| 91 | rank = np.array(rank[1::2]).flatten()
|
|---|
| 92 | print(rank)
|
|---|
| 93 | rankTemperature = zip(rank, self._T)
|
|---|
| 94 | rankTemperature.sort()
|
|---|
| 95 | print self._myPTRank, rankTemperature[self._myReplica][1]
|
|---|
| 96 | # self._myUniverse.setTemperature(rankTemperature[self._myReplica][1])
|
|---|
| 97 | else:
|
|---|
| 98 | random.random() # Do one random number to stay in sync with the master process for the same replica.
|
|---|
| 99 |
|
|---|
| 100 |
|
|---|
| 101 |
|
|---|
| 102 |
|
|---|
| 103 | if __name__ == "__main__":
|
|---|
| 104 |
|
|---|
| 105 | n = int(sys.argv[1])
|
|---|
| 106 | T = [250, 275, 300, 330, 370, 420]
|
|---|
| 107 |
|
|---|
| 108 | myPT = ParallelTempering(T, [])
|
|---|
| 109 | myPT.exchange()
|
|---|
| 110 | myPT.exchange()
|
|---|
| 111 |
|
|---|
| 112 |
|
|---|
| 113 |
|
|---|
| 114 |
|
|---|
| 115 |
|
|---|
| 116 |
|
|---|
| 117 |
|
|---|
