Weird fluctuations in Ising model

[mastrofoffi]

Not 100% sure this thread belongs in this section, sorry if it's out of place.

I was trying to run a simulation of a 2D Ising model; i thought everything was going fine until I started to look at the numerical results and I noticed that I get some wild fluctuations in my results; as an example, here are the values i get for final energy and magnetization (per site) after 10 different 10000-step runs on a 128x128 lattice at temperature T = 0.1 K(critical temperature is at 2.27 K), where the initial spin configuration is chosen randomly(+1 or -1 is assigned to each node with equal probability):
$$ (E, |M|) = (-2, 1); (-1.95, 0.52); (-2, 1); (-1.95, 0.21); (-2, 1); (-1.95, 0.02); (-2, 1); (-1.95; 0.24); (-2, 1); (-2; 1) $$
With a temperature so low I would expect all $|M|$ to be exactly 1, or at least that fluctuations like these would be way less probable than 4/10, so I guess there must be some error in my code. Am I wrong in thinking this?
I would be very grateful if someone could take a look at my code and maybe point out where the error is, cause it seems like I can't find it.

Here is the function used to update the lattice(it's Metropolis algorithm):

void update(graph *pLattice, int *s, double *prob, int *pE, int *pM){
  /* pLattice is the pointer to the graph/lattice; s is the pointer
   * to the array containing the spin value for each site; prob is the pointer to
   * the lookup table used to map differences in energy to their probability according
   * to Boltzmann distribution; pE and pM point respectively to the energy and 
   * magnetization values of the lattice. */
  int k, site, *pNeigh, siteConn, sum;
  for(site = 0; site < pLattice->numNodes; site++){ // loop over all nodes
    pNeigh = pLattice->neigh[site]; // = neigh[site][0]
    siteConn = pLattice->conn[site]; // = 4
    sum = 0;
    for(k = 0; k < siteConn; k++, pNeigh++){
      sum += s[*pNeigh];  // = s[neigh[site][k]]
    }
    sum *= s[site]; // this is now the energy of site
    if(sum <= 0 || drand48() < prob[sum]){
      /* this is legit because sum = half the variation of energy due to the flip of s[site]
       * and 0 < prob[sum] = exp(-2.0 * sum / T) < 1; (T=temperature) */
      s[site] = -s[site];
      *pE += 2*sum;
      *pM += 2*s[site];
    }
  }
  return;
}

"graph" is basically an adjacency list, defined as

typedef struct {
  /* conn points to an array of size numNodes, whose items are the number of
   * connections of each node(i'm using periodic bc so conn[i] = 4 for every value of i).
   * neigh points to a 2d array of dimensions numNodes*numLinks,
   * where the neighbours of each node are stored. */
  int numNodes, numLinks, *conn, **neigh;
} graph;

The main function, besides declaration of variables and error checking, is basically just

  E = energy(pLattice, s);
  M = magnetization(pLattice, s);
  for(i=0;i<R;i++){
    update(pLattice, s, probLUT, &E, &M);    
  }
  fprintf(stderr, "E = %f; M = %f\n", (double)E/N, (double)M/N);

where N is the number of nodes.

The difference in energy value between "correct" and "weird" runs is small, so when I average over lot of runs for each T, and for many values of T, the plot EvsT
looks not too different from the expected curve; the plot of |M|vsT still shows the correct behaviour, but instead of getting constant |M|=1 for $T<T_c$ i get values oscillating around |M|=0.8. I tried running for different times, different lattice size, different initial conditions, with no different results.
I'll post full code, graphs or whatever if needed. Thanks.

 

[mastrofoffi]

Updating just in case anyone happens to come by and help me out. I created some animated plots for the 'weird' results and I found that they correspond to situations where two connected clusters(one for each spin) arise and reach some sort of equilibrium configuration where they are just translated (always) in the $-x$ direction through the periodic boundary. Here a few pics of a 64x64 lattice taken at times 850,851,852,856:

Basically it evolves for some time, then stabilizes on a configuration like this which is mantained for the rest of the simulation. This makes no sense at all to me: why would something like this happen? and how? I also tried rewriting the program all over, without looking at the original one, but I get the exact same thing. I guess this must be related to the periodic boundaries, so I'll try to implement free boundaries and see what happens.