diff --git a/src/rcb.c b/src/rcb.c
index 5defa8c5..802c5ced 100644
--- a/src/rcb.c
+++ b/src/rcb.c
@@ -143,9 +143,7 @@ int rcb(struct array *elements, size_t unit_size, int ndim, struct comm *ci,
   struct comm c, t;
   comm_dup(&c, ci);
 
-  int size = c.np;
-  int rank = c.id;
-
+  int size = c.np, rank = c.id;
   while (size > 1) {
     rcb_level(elements, unit_size, ndim, &c, bfr);
 
diff --git a/src/rib.c b/src/rib.c
index 1aac9031..889848b6 100644
--- a/src/rib.c
+++ b/src/rib.c
@@ -2,18 +2,14 @@
 #include "sort.h"
 
 extern int power_serial(double *y, int N, double *A, int verbose);
+extern int inv_power_serial(double *y, uint N, double *A, int verbose);
 
-static void get_rib_axis(char *elems, uint nel, size_t unit_size, int ndim,
+static void get_rib_proj(char *elems, uint nel, size_t unit_size, int ndim,
                          struct comm *c) {
-  double avg[3];
-  avg[0] = avg[1] = avg[2] = 0.0;
-  struct rcb_element *elem;
-  uint i;
-  for (i = 0; i < nel; i++) {
-    elem = (struct rcb_element *)(elems + i * unit_size);
-    avg[0] += elem->coord[0];
-    avg[1] += elem->coord[1];
-    avg[2] += elem->coord[2];
+  double avg[3] = {0, 0, 0};
+  for (uint i = 0; i < nel; i++) {
+    struct rcb_element *ei = (struct rcb_element *)(elems + i * unit_size);
+    avg[0] += ei->coord[0], avg[1] += ei->coord[1], avg[2] += ei->coord[2];
   }
   slong nelg = nel;
 
@@ -23,37 +19,44 @@ static void get_rib_axis(char *elems, uint nel, size_t unit_size, int ndim,
     comm_allreduce(c, gs_long, gs_add, &nelg, 1, buf);
   }
 
-  avg[0] /= nelg;
-  avg[1] /= nelg;
-  avg[2] /= nelg;
+  avg[0] /= nelg, avg[1] /= nelg, avg[2] /= nelg;
 
-  double I[3][3];
-  for (i = 0; i < 3; i++)
-    I[i][0] = I[i][1] = I[i][2] = 0.0;
+  double I[9];
+  for (unsigned i = 0; i < 9; i++)
+    I[i] = 0;
 
   double x, y, z;
-  for (i = 0; i < nel; i++) {
-    elem = (struct rcb_element *)(elems + i * unit_size);
-    x = elem->coord[0] - avg[0];
-    y = elem->coord[1] - avg[1];
-    z = elem->coord[2] - avg[2];
-    I[0][0] += x * x, I[0][1] += x * y, I[0][2] += x * z;
-    I[1][0] += y * x, I[1][1] += y * y, I[1][2] += y * z;
-    I[2][0] += z * x, I[2][1] += z * y, I[2][2] += z * z;
+  for (uint i = 0; i < nel; i++) {
+    struct rcb_element *ei = (struct rcb_element *)(elems + i * unit_size);
+    x = ei->coord[0] - avg[0];
+    y = ei->coord[1] - avg[1];
+    z = ei->coord[2] - avg[2];
+    I[0] += x * x, I[1] += x * y, I[2] += x * z;
+    I[3] += y * x, I[4] += y * y, I[5] += y * z;
+    I[6] += z * x, I[7] += z * y, I[8] += z * z;
   }
 
   if (c != NULL)
     comm_allreduce(c, gs_double, gs_add, I, 9, buf);
 
-  double ev[3];                           // ev[2] = 0 if 2D
-  power_serial(ev, ndim, (double *)I, 0); // FIXME: 2D does not work
+  // FIXME: 2D does not work
+  double ev[3];
+  power_serial(ev, ndim, (double *)I, 0);
 
-  for (i = 0; i < nel; i++) {
-    elem = (struct rcb_element *)(elems + i * unit_size);
-    x = elem->coord[0] - avg[0];
-    y = elem->coord[1] - avg[1];
-    z = elem->coord[2] - avg[2];
-    elem->fiedler = x * ev[0] + y * ev[1] + z * ev[2];
+  double norm = 0;
+  for (unsigned i = 0; i < ndim; i++)
+    norm += ev[i] * ev[i];
+  norm = sqrt(norm);
+
+  for (unsigned i = 0; i < ndim; i++)
+    ev[i] /= norm;
+
+  for (uint i = 0; i < nel; i++) {
+    struct rcb_element *ei = (struct rcb_element *)(elems + i * unit_size);
+    x = ei->coord[0] - avg[0];
+    y = ei->coord[1] - avg[1];
+    z = ei->coord[2] - avg[2];
+    ei->fiedler = x * ev[0] + y * ev[1] + z * ev[2];
   }
 }
 
@@ -64,7 +67,7 @@ void rib_local(struct array *a, size_t unit_size, uint start, uint end,
     return;
 
   char *st = (char *)a->ptr + unit_size * start;
-  get_rib_axis(st, size, unit_size, ndim, NULL);
+  get_rib_proj(st, size, unit_size, ndim, NULL);
 
   if (unit_size == sizeof(struct rcb_element))
     sarray_sort(struct rcb_element, st, size, fiedler, 3, buf);
@@ -81,7 +84,7 @@ static int rib_level(struct array *a, size_t unit_size, int ndim,
   if (c->np == 1)
     return 0;
 
-  get_rib_axis((char *)a->ptr, a->n, unit_size, ndim, c);
+  get_rib_proj((char *)a->ptr, a->n, unit_size, ndim, c);
 
   if (unit_size == sizeof(struct rcb_element))
     parallel_sort(struct rcb_element, a, fiedler, gs_double, 0, 1, c, bfr);
@@ -93,26 +96,23 @@ static int rib_level(struct array *a, size_t unit_size, int ndim,
 
 int rib(struct array *elements, size_t unit_size, int ndim, struct comm *ci,
         buffer *bfr) {
-  struct comm c;
+  struct comm c, t;
   comm_dup(&c, ci);
 
-  int size = c.np;
-  int rank = c.id;
-
+  int size = c.np, rank = c.id;
   while (size > 1) {
     rib_level(elements, unit_size, ndim, &c, bfr);
 
-    int p = (size + 1) / 2;
-    int bin = (rank >= p);
+    int bin = 1;
+    if (rank < (size + 1) / 2)
+      bin = 0;
 
-    MPI_Comm comm_rib;
-    MPI_Comm_split(c.c, bin, rank, &comm_rib);
+    comm_split(&c, bin, rank, &t);
     comm_free(&c);
-    comm_init(&c, comm_rib);
-    MPI_Comm_free(&comm_rib);
+    comm_dup(&c, &t);
+    comm_free(&t);
 
-    size = c.np;
-    rank = c.id;
+    size = c.np, rank = c.id;
   }
   comm_free(&c);
 
diff --git a/src/rsb.c b/src/rsb.c
index ea6bce8a..2ec28d9f 100644
--- a/src/rsb.c
+++ b/src/rsb.c
@@ -10,6 +10,8 @@ extern int rsb(struct array *elements, int nv, int check,
                parrsb_options *options, struct comm *gc, buffer *bfr);
 extern int rcb(struct array *elements, size_t unit_size, int ndim,
                struct comm *ci, buffer *bfr);
+extern int rib(struct array *elements, size_t unit_size, int ndim,
+               struct comm *ci, buffer *bfr);
 
 parrsb_options parrsb_default_options = {
     // General options
@@ -56,7 +58,7 @@ static void update_options(parrsb_options *options) {
   UPDATE_OPTION(rsb_mg_grammian, "PARRSB_RSB_MG_GRAMMIAN", 1);
   UPDATE_OPTION(rsb_mg_factor, "PARRSB_RSB_MG_FACTOR", 1);
   UPDATE_OPTION(rsb_mg_sagg, "PARRSB_RSB_MG_SMOOTH_AGGREGATION", 1);
-  if (options->verbose_level == 0) 
+  if (options->verbose_level == 0)
     options->profile_level = 0;
 }
 
@@ -99,11 +101,11 @@ static size_t load_balance(struct array *elist, uint nel, int nv, double *coord,
     unit_size = sizeof(struct rsb_element);
 
   array_init_(elist, nel, unit_size, __FILE__, __LINE__);
+  elist->n = 0;
 
+  int ndim = (nv == 8) ? 3 : 2;
   struct rcb_element *pe = (struct rcb_element *)calloc(1, unit_size);
   pe->origin = c->id;
-
-  int ndim = (nv == 8) ? 3 : 2;
   for (uint e = 0; e < nel; ++e) {
     slong eg = pe->globalId = start + e + 1;
     if (eg <= lower)
@@ -200,10 +202,12 @@ int parrsb_part_mesh(int *part, int *seq, long long *vtx, double *coord,
     slong nelg = out[1][0];
 
     if (ca.np > nelg) {
-      if (ca.id == 0)
-        printf("Total number of elements is smaller than the "
-               "number of processors.\n"
-               "Run with smaller number of processors.\n");
+      if (ca.id == 0) {
+        fprintf(stderr, "Total number of elements is smaller than the "
+                        "number of processors.\n"
+                        "Run with smaller number of processors.\n");
+        fflush(stderr);
+      }
       return 1;
     }