diff --git a/src/bitbool.h b/src/bitbool.h
index 695b2a0..a408726 100644
--- a/src/bitbool.h
+++ b/src/bitbool.h
@@ -38,11 +38,11 @@ static const cmph_uint8 valuemask[] = { 0xfc, 0xf3, 0xcf, 0x3f};
  * 
  * UNSETBIT(array, i) is a macro that sets 0 to an 1-bit integer stored in an array.
  */
-#define UNSETBIT(array, i) (array[i >> 3] &= (~(bitmask[i & 0x00000007])))
+#define UNSETBIT(array, i) (array[i >> 3] ^= ((bitmask[i & 0x00000007])))
 
 //#define GETBIT(array, i) (array[(i) / 8] & bitmask[(i) % 8])
 //#define SETBIT(array, i) (array[(i) / 8] |= bitmask[(i) % 8])
-//#define UNSETBIT(array, i) (array[(i) / 8] &= (~(bitmask[(i) % 8])))
+//#define UNSETBIT(array, i) (array[(i) / 8] ^= ((bitmask[(i) % 8])))
 
 
 /** \def SETVALUE1(array, i, v)
@@ -99,6 +99,15 @@ static const cmph_uint8 valuemask[] = { 0xfc, 0xf3, 0xcf, 0x3f};
  */
 #define GETBIT32(array, i) (array[i >> 5] & bitmask32[i & 0x0000001f])
 
+/** \def UNSETBIT32(array, i)
+ *  \brief set 0 to an 1-bit integer stored in an array of 32-bit words. 
+ *  \param array to store 1-bit integer values. The entries ar 32-bit words
+ *  \param i is the index in array to set the the bit to 0
+ * 
+ * UNSETBIT32(array, i) is a macro that sets 0 to an 1-bit integer stored in an array of 32-bit words.
+ */
+#define UNSETBIT32(array, i) (array[i >> 5] ^= ((bitmask32[i & 0x0000001f])))
+
 #define BITS_TABLE_SIZE(n, bits_length) ((n * bits_length + 31) >> 5)
 
 static inline void set_bits_value(cmph_uint32 * bits_table, cmph_uint32 index, cmph_uint32 bits_string, 
diff --git a/src/chd.c b/src/chd.c
index 72523c0..eaf4909 100644
--- a/src/chd.c
+++ b/src/chd.c
@@ -9,7 +9,7 @@
 #include "cmph_structs.h"
 #include "chd_structs.h"
 #include "chd.h"
-
+#include "bitbool.h"
 //#define DEBUG
 #include "debug.h"
 
@@ -78,7 +78,7 @@ cmph_t *chd_new(cmph_config_t *mph, double c)
 
 	register cmph_uint32 i, idx, nkeys, nvals, nbins;
 	cmph_uint32 * vals_table = NULL;
-	register cmph_uint8 * occup_table = NULL;
+	register cmph_uint32 * occup_table = NULL;
 	#ifdef CMPH_TIMING
 	double construction_time_begin = 0.0;
 	double construction_time = 0.0;
@@ -123,11 +123,11 @@ cmph_t *chd_new(cmph_config_t *mph, double c)
 	nvals =  nbins - nkeys; 
 	
 	vals_table = (cmph_uint32 *)calloc(nvals, sizeof(cmph_uint32));
-	occup_table = chd_ph->occup_table;
+	occup_table = (cmph_uint32 *)chd_ph->occup_table;
 	
 	for(i = 0, idx = 0; i < nbins; i++)
 	{
-		if(occup_table[i] == 0)
+		if(!GETBIT32(occup_table, i))
 		{
 			vals_table[idx++] = i;
 		}
diff --git a/src/chd_ph.c b/src/chd_ph.c
index 68fe6c7..71acc4b 100644
--- a/src/chd_ph.c
+++ b/src/chd_ph.c
@@ -10,6 +10,8 @@
 #include "chd_structs_ph.h"
 #include "chd_ph.h"
 #include"miller_rabin.h"
+#include"bitbool.h"
+
 
 //#define DEBUG
 #include "debug.h"
@@ -19,28 +21,47 @@
 #define NO_ELEMENT UINT_MAX
 #endif
 
-// struct to represents the buckets items
+// struct used to represent items at mapping, ordering and searching phases
 struct _chd_ph_item_t
 {
 	cmph_uint32 f;
 	cmph_uint32 h;
-	struct _chd_ph_item_t * next;
 };
 typedef struct _chd_ph_item_t chd_ph_item_t;
 
+// struct to represent the items at mapping phase only. 
+struct _chd_ph_map_item_t
+{
+	cmph_uint32 f;
+	cmph_uint32 h;
+	cmph_uint32 bucket_num;
+};
+typedef struct _chd_ph_map_item_t chd_ph_map_item_t;
 
 // struct to represent a bucket
 struct _chd_ph_bucket_t
 {
-	cmph_uint32 size;
-	chd_ph_item_t * items_list;
-	cmph_uint32 next_in_list;
+	cmph_uint32 items_list; // offset
+	union
+	{
+		cmph_uint32 size;
+		cmph_uint32 bucket_id;
+	};
 };
+
 typedef struct _chd_ph_bucket_t chd_ph_bucket_t;
 
+struct _chd_ph_sorted_list_t
+{
+	cmph_uint32 buckets_list;
+	cmph_uint32 size;
+};
+
+typedef struct _chd_ph_sorted_list_t chd_ph_sorted_list_t;
+
+
 static inline chd_ph_bucket_t * chd_ph_bucket_new(cmph_uint32 nbuckets);
 static inline void chd_ph_bucket_clean(chd_ph_bucket_t * buckets, cmph_uint32 nbuckets);
-static inline cmph_uint8 chd_ph_bucket_insert(chd_ph_bucket_t * buckets, cmph_uint32 nbuckets, cmph_uint32 g, chd_ph_item_t * item);
 static inline void chd_ph_bucket_destroy(chd_ph_bucket_t * buckets);
 
 chd_ph_bucket_t * chd_ph_bucket_new(cmph_uint32 nbuckets)
@@ -54,40 +75,31 @@ void chd_ph_bucket_clean(chd_ph_bucket_t * buckets, cmph_uint32 nbuckets)
 	register cmph_uint32 i = 0;
 	assert(buckets);
 	for(i = 0; i < nbuckets; i++)
-	{
 		buckets[i].size = 0;
-		buckets[i].items_list = 0;
-		buckets[i].next_in_list = NO_ELEMENT;
-	};
 }
-
-cmph_uint8 chd_ph_bucket_insert(chd_ph_bucket_t * buckets, cmph_uint32 nbuckets, cmph_uint32 g, chd_ph_item_t * item)
+cmph_uint8 chd_ph_bucket_insert(chd_ph_bucket_t * buckets,chd_ph_map_item_t * map_items, chd_ph_item_t * items,
+				cmph_uint32 nbuckets,cmph_uint32 item_idx)
 {
-	chd_ph_item_t * item1, * prior_item1;
-	item1 = buckets[g].items_list;
-	prior_item1 = 0;
-	while(item1 != 0 && (item1->f < item->f || (item1->f == item->f && item1->h < item->h)) )
+	register cmph_uint32 i = 0;
+	register chd_ph_item_t * tmp_item;
+	register chd_ph_map_item_t * tmp_map_item = map_items + item_idx;
+	register chd_ph_bucket_t * bucket = buckets + tmp_map_item->bucket_num;
+	tmp_item = items + bucket->items_list;
+	
+	for(i = 0; i < bucket->size; i++)
 	{
-		prior_item1 = item1;
-		item1 = item1->next;
+		if(tmp_item->f == tmp_map_item->f && tmp_item->h == tmp_map_item->h)
+		{
+			DEBUGP("Item not added\n");
+			return 0;
+		};
+		tmp_item++;
 	};
-	
-	if(item1 != 0 && item1->f == item->f && item1->h == item->h)
-	{
-		DEBUGP("Item not added\n");
-		return 0;
-	};
-	item->next = item1;
-	if(prior_item1 == 0)
-		buckets[g].items_list = item;
-	else
-		prior_item1->next = item;
-	
-	buckets[g].size++;
-	
+	tmp_item->f = tmp_map_item->f;
+	tmp_item->h = tmp_map_item->h;
+	bucket->size++;
 	return 1;
-}
-
+};
 void chd_ph_bucket_destroy(chd_ph_bucket_t * buckets)
 {
     free(buckets);
@@ -96,10 +108,11 @@ void chd_ph_bucket_destroy(chd_ph_bucket_t * buckets)
 static inline cmph_uint8 chd_ph_mapping(cmph_config_t *mph, chd_ph_bucket_t * buckets, chd_ph_item_t * items, 
 					cmph_uint32 *max_bucket_size);
 
-static inline cmph_uint32 * chd_ph_ordering(chd_ph_bucket_t * buckets, cmph_uint32 nbuckets, cmph_uint32 max_bucket_size);
+static chd_ph_sorted_list_t * chd_ph_ordering(chd_ph_bucket_t ** _buckets,chd_ph_item_t ** items,
+				cmph_uint32 nbuckets,cmph_uint32 nitems, cmph_uint32 max_bucket_size);
 
-static inline cmph_uint8 chd_ph_searching(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, cmph_uint32 max_bucket_size,
-                                          cmph_uint32 *sorted_lists, cmph_uint32 max_probes, cmph_uint32 * disp_table);
+static cmph_uint8 chd_ph_searching(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, chd_ph_item_t *items ,
+	cmph_uint32 max_bucket_size, chd_ph_sorted_list_t *sorted_lists, cmph_uint32 max_probes, cmph_uint32 * disp_table);
 
 static inline double chd_ph_space_lower_bound(cmph_uint32 _n, cmph_uint32 _r)
 {
@@ -207,7 +220,8 @@ cmph_uint8 chd_ph_mapping(cmph_config_t *mph, chd_ph_bucket_t * buckets, chd_ph_
 	chd_ph_config_data_t *chd_ph = (chd_ph_config_data_t *)mph->data;
 	char * key = NULL;
 	cmph_uint32 keylen = 0;
-	chd_ph_item_t * item;
+	chd_ph_map_item_t * map_item;
+	chd_ph_map_item_t * map_items = malloc(chd_ph->m*sizeof(chd_ph_map_item_t));
 	register cmph_uint32 mapping_iterations = 1000;
 	*max_bucket_size = 0;
 	while(1)
@@ -225,33 +239,40 @@ cmph_uint8 chd_ph_mapping(cmph_config_t *mph, chd_ph_bucket_t * buckets, chd_ph_
 			mph->key_source->read(mph->key_source->data, &key, &keylen);		
 			hash_vector(chd_ph->hl, key, keylen, hl);
 			
-			item = (items + i);
+			map_item = (map_items + i);
 
 			g = hl[0] % chd_ph->nbuckets;
-			item->f = hl[1] % chd_ph->n;
-			item->h = hl[2] % (chd_ph->n - 1) + 1;
-			
+			map_item->f = hl[1] % chd_ph->n;
+			map_item->h = hl[2] % (chd_ph->n - 1) + 1;
+			map_item->bucket_num=g;
 			mph->key_source->dispose(mph->key_source->data, key, keylen);		
-			
 // 			if(buckets[g].size == (chd_ph->keys_per_bucket << 2))
 // 			{
 // 				DEBUGP("BUCKET = %u -- SIZE = %u -- MAXIMUM SIZE = %u\n", g, buckets[g].size, (chd_ph->keys_per_bucket << 2));
 // 				goto error;
 // 			}
-			
-			if(!chd_ph_bucket_insert(buckets, chd_ph->nbuckets, g, item))
-			{
-				break;
-			}
-			
+			buckets[g].size++;
 			if(buckets[g].size > *max_bucket_size)
 			{
 				  *max_bucket_size = buckets[g].size;
 			}
 		}
-		
+		buckets[0].items_list = 0;
+		for(i = 1; i < chd_ph->nbuckets; i++)
+		{
+			buckets[i].items_list = buckets[i-1].items_list + buckets[i - 1].size;
+			buckets[i - 1].size = 0;
+		};
+		buckets[i - 1].size = 0;
+		for(i = 0; i < chd_ph->m; i++)
+		{
+			map_item = (map_items + i);
+			if(!chd_ph_bucket_insert(buckets, map_items, items, chd_ph->nbuckets, i))
+				break;
+		}
 		if(i == chd_ph->m)
 		{
+			free(map_items);
 			return 1; // SUCCESS
 		}
 		
@@ -261,77 +282,163 @@ cmph_uint8 chd_ph_mapping(cmph_config_t *mph, chd_ph_bucket_t * buckets, chd_ph_
 		}
 	}
 error:
+	free(map_items);
 	hash_state_destroy(chd_ph->hl);
 	chd_ph->hl = NULL;
 	return 0; // FAILURE
-
 }
 
-cmph_uint32 * chd_ph_ordering(chd_ph_bucket_t * buckets, cmph_uint32 nbuckets, cmph_uint32 max_bucket_size)
+chd_ph_sorted_list_t * chd_ph_ordering(chd_ph_bucket_t ** _buckets, chd_ph_item_t ** _items,
+	cmph_uint32 nbuckets, cmph_uint32 nitems, cmph_uint32 max_bucket_size)
 {
-	cmph_uint32 * sorted_lists = (cmph_uint32 *) calloc(max_bucket_size + 1, sizeof(cmph_uint32));
-	register cmph_uint32 i, size;
+	chd_ph_sorted_list_t * sorted_lists = (chd_ph_sorted_list_t *) calloc(max_bucket_size + 1, sizeof(chd_ph_sorted_list_t));
+	
+	chd_ph_bucket_t * input_buckets = (*_buckets);
+	chd_ph_bucket_t * output_buckets;
+	chd_ph_item_t * input_items = (*_items);
+	chd_ph_item_t * output_items;
+	register cmph_uint32 i, j, bucket_size, position, position2;
+// 	cmph_uint32 non_empty_buckets;
 	DEBUGP("MAX BUCKET SIZE = %u\n", max_bucket_size);
-	for(i = 0; i <= max_bucket_size; i++)
-	{
-		sorted_lists[i] = NO_ELEMENT;
-	}
+	// Determine size of each list of buckets
 	for(i = 0; i < nbuckets; i++)
 	{
-		size = buckets[i].size;
-		if(size == 0)
+		bucket_size = input_buckets[i].size;
+		if(bucket_size == 0)
 			continue;
-		buckets[i].next_in_list = sorted_lists[size];
-		sorted_lists[size] = i;
+		sorted_lists[bucket_size].size++;
 	};
-
+	sorted_lists[1].buckets_list = 0;
+	// Determine final position of list of buckets into the contiguous array that will store all the buckets
+	for(i = 2; i <= max_bucket_size; i++)
+	{
+		sorted_lists[i].buckets_list = sorted_lists[i-1].buckets_list + sorted_lists[i-1].size;
+		sorted_lists[i-1].size = 0;
+	};
+	sorted_lists[i-1].size = 0;
+	// Store the buckets in a new array which is sorted by bucket sizes
+	output_buckets = calloc(nbuckets, sizeof(chd_ph_bucket_t)); // everything is initialized with zero
+//  	non_empty_buckets = nbuckets;
+	
+	for(i = 0; i < nbuckets; i++)
+	{
+		bucket_size = input_buckets[i].size;
+		if(bucket_size == 0)
+		{
+// 			non_empty_buckets--;
+			continue;
+		};
+		position = sorted_lists[bucket_size].buckets_list + sorted_lists[bucket_size].size;
+		output_buckets[position].bucket_id = i;
+		output_buckets[position].items_list = input_buckets[i].items_list;
+		sorted_lists[bucket_size].size++;
+	};
+/*	for(i = non_empty_buckets; i < nbuckets; i++)
+		output_buckets[i].size=0;*/
+	// Return the buckets sorted in new order and free the old buckets sorted in old order
+	free(input_buckets);
+	(*_buckets) = output_buckets;
+	
+	
+	// Store the items according to the new order of buckets.
+	output_items = (chd_ph_item_t*)calloc(nitems, sizeof(chd_ph_item_t));
+	position = 0;
+	i = 0;
+	for(bucket_size = 1; bucket_size <= max_bucket_size; bucket_size++)
+	{
+		for(i = sorted_lists[bucket_size].buckets_list; i < sorted_lists[bucket_size].size + sorted_lists[bucket_size].buckets_list; i++)
+		{
+			position2 = output_buckets[i].items_list;
+			output_buckets[i].items_list = position;
+			for(j = 0; j < bucket_size; j++)
+			{
+				output_items[position].f = input_items[position2].f;
+				output_items[position].h = input_items[position2].h;
+				position++;
+				position2++;
+			};
+		};
+	};
+	//Return the items sorted in new order and free the old items sorted in old order
+	free(input_items);
+	(*_items) = output_items;
 	return sorted_lists;
-}
+};
 
-static inline cmph_uint8 place_bucket_probe(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, cmph_uint32 probe0_num, 
-					    cmph_uint32 probe1_num, cmph_uint32 bucket_num)
+static inline cmph_uint8 place_bucket_probe(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets,
+					    chd_ph_item_t *items, cmph_uint32 probe0_num, cmph_uint32 probe1_num,
+					    cmph_uint32 bucket_num, cmph_uint32 size)
 {
 	register cmph_uint32 i;
-	register cmph_uint32 size = buckets[bucket_num].size;
 	register chd_ph_item_t * item;
 	register cmph_uint32 position;
 
-	item = buckets[bucket_num].items_list;
+	item = items + buckets[bucket_num].items_list;
 	// try place bucket with probe_num
-	for(i = 0; i < size; i++) // placement
+	if(chd_ph->keys_per_bin > 1)
 	{
-		position = (item->f + ((cmph_uint64)item->h)*probe0_num + probe1_num) % chd_ph->n;
-		
-		if(chd_ph->occup_table[position] >= chd_ph->keys_per_bin)
+		for(i = 0; i < size; i++) // placement
 		{
-			break;
-		}
-		(chd_ph->occup_table[position])++;
-		
-		item = item->next;
-	};
-	
-	if(i != size) // Undo the placement
-	{
-		item = buckets[bucket_num].items_list;
-		while(1)
-		{
-			if(i == 0)
+			position = (item->f + ((cmph_uint64)item->h)*probe0_num + probe1_num) % chd_ph->n;
+			if(chd_ph->occup_table[position] >= chd_ph->keys_per_bin)
 			{
 				break;
 			}
-			position = (item->f + ((cmph_uint64 )item->h) * probe0_num + probe1_num) % chd_ph->n;
-			(chd_ph->occup_table[position])--;
-			item = item->next;
-			i--;
+			(chd_ph->occup_table[position])++;
+			item++;
+		};
+	} else
+	{
+		for(i = 0; i < size; i++) // placement
+		{
+			position = (item->f + ((cmph_uint64)item->h)*probe0_num + probe1_num) % chd_ph->n;
+			if(GETBIT32(((cmph_uint32 *)chd_ph->occup_table), position))
+			{
+				break;
+			}
+			SETBIT32(((cmph_uint32*)chd_ph->occup_table), position);
+			item++;
+		};
+	};
+	if(i != size) // Undo the placement
+	{
+		item = items + buckets[bucket_num].items_list;
+		if(chd_ph->keys_per_bin > 1)
+		{
+			while(1)
+			{
+				if(i == 0)
+				{
+					break;
+				}
+				position = (item->f + ((cmph_uint64 )item->h) * probe0_num + probe1_num) % chd_ph->n;
+				(chd_ph->occup_table[position])--;
+				item++;
+				i--;
+			};
+		} else
+		{
+			while(1)
+			{
+				if(i == 0)
+				{
+					break;
+				}
+				position = (item->f + ((cmph_uint64 )item->h) * probe0_num + probe1_num) % chd_ph->n;
+				UNSETBIT32(((cmph_uint32*)chd_ph->occup_table), position);
+				
+// 				([position/32]^=(1<<(position%32));
+				item++;
+				i--;
+			};
 		};
 		return 0;
-	};	
+	} 	
 	return 1;
 };
 
-static inline cmph_uint8 place_bucket(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, cmph_uint32 max_probes, 
-                                      cmph_uint32 * disp_table, cmph_uint32 bucket_num)
+static inline cmph_uint8 place_bucket(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, chd_ph_item_t * items, cmph_uint32 max_probes, 
+                                      cmph_uint32 * disp_table, cmph_uint32 bucket_num, cmph_uint32 size)
 				      
 {
 	register cmph_uint32 probe0_num, probe1_num, probe_num;
@@ -341,9 +448,9 @@ static inline cmph_uint8 place_bucket(chd_ph_config_data_t *chd_ph, chd_ph_bucke
 	
 	while(1)
 	{
-		if(place_bucket_probe(chd_ph, buckets, probe0_num, probe1_num, bucket_num))
+		if(place_bucket_probe(chd_ph, buckets, items, probe0_num, probe1_num, bucket_num,size))
 		{
-			disp_table[bucket_num] = probe0_num + probe1_num * chd_ph->n;
+			disp_table[buckets[bucket_num].bucket_id] = probe0_num + probe1_num * chd_ph->n;
 			return 1;
 		}
 		probe0_num++;
@@ -361,68 +468,76 @@ static inline cmph_uint8 place_bucket(chd_ph_config_data_t *chd_ph, chd_ph_bucke
 	return 0;
 };
 
-static inline cmph_uint8 place_buckets1(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, cmph_uint32 max_bucket_size,
-                                        cmph_uint32 *sorted_lists, cmph_uint32 max_probes, cmph_uint32 * disp_table)
+static inline cmph_uint8 place_buckets1(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t * buckets, chd_ph_item_t *items,
+					cmph_uint32 max_bucket_size, chd_ph_sorted_list_t *sorted_lists, cmph_uint32 max_probes, 
+					cmph_uint32 * disp_table)
 {
 	register cmph_uint32 i = 0;
 	register cmph_uint32 curr_bucket = 0;
 
 	for(i = max_bucket_size; i > 0; i--)
 	{
-		curr_bucket = sorted_lists[i];
-		while(curr_bucket != NO_ELEMENT)
+		curr_bucket = sorted_lists[i].buckets_list;
+		while(curr_bucket < sorted_lists[i].size + sorted_lists[i].buckets_list)
 		{
-			if(!place_bucket(chd_ph, buckets, max_probes, disp_table, curr_bucket))
+			if(!place_bucket(chd_ph, buckets, items, max_probes, disp_table, curr_bucket, i))
 			{
 				return 0;
 			}
-			curr_bucket = buckets[curr_bucket].next_in_list;
+			curr_bucket++;
 		};
-	
 	};
 	return 1;
 };
 
-static inline cmph_uint8 place_buckets2(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, cmph_uint32 max_bucket_size,
-                                        cmph_uint32 *sorted_lists, cmph_uint32 max_probes, cmph_uint32 * disp_table)
+static inline cmph_uint8 place_buckets2(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, chd_ph_item_t * items, 
+					cmph_uint32 max_bucket_size, chd_ph_sorted_list_t *sorted_lists, cmph_uint32 max_probes, 
+					cmph_uint32 * disp_table)
 {
-	register cmph_uint32 i;
-	register cmph_uint32 curr_bucket, prev_bucket;
+	register cmph_uint32 i,j, non_placed_bucket;
+	register cmph_uint32 curr_bucket;
 	register cmph_uint32 probe_num, probe0_num, probe1_num;
+	cmph_uint32 sorted_list_size;
+#ifdef DEBUG
+	cmph_uint32 items_list;
+	cmph_uint32 bucket_id;
+#endif
 	DEBUGP("USING HEURISTIC TO PLACE BUCKETS\n");
 	for(i = max_bucket_size; i > 0; i--)
 	{
 		probe_num = 0;
 		probe0_num = 0;
 		probe1_num = 0;
-		while(sorted_lists[i] != NO_ELEMENT)
+		sorted_list_size = sorted_lists[i].size;
+		while(sorted_lists[i].size != 0)
 		{
-			prev_bucket = NO_ELEMENT;
-			curr_bucket = sorted_lists[i];
-			while(curr_bucket != NO_ELEMENT)
+			curr_bucket = sorted_lists[i].buckets_list;
+			for(j = 0, non_placed_bucket = 0; j < sorted_lists[i].size; j++)
 			{
 				// if bucket is successfully placed remove it from list
-				if(place_bucket_probe(chd_ph, buckets, probe0_num, probe1_num, curr_bucket))
+				if(place_bucket_probe(chd_ph, buckets, items, probe0_num, probe1_num, curr_bucket, i))
 				{	
-					disp_table[curr_bucket] = probe0_num + probe1_num * chd_ph->n;
+					disp_table[buckets[curr_bucket].bucket_id] = probe0_num + probe1_num * chd_ph->n;
 // 					DEBUGP("BUCKET %u PLACED --- DISPLACEMENT = %u\n", curr_bucket, disp_table[curr_bucket]);
-					if(prev_bucket == NO_ELEMENT)
-					{
-						sorted_lists[i] = buckets[curr_bucket].next_in_list;
-					}
-					else
-					{
-						buckets[prev_bucket].next_in_list = buckets[curr_bucket].next_in_list;
-					}
-						
 				} 
 				else
 				{
 // 					DEBUGP("BUCKET %u NOT PLACED\n", curr_bucket);
-					prev_bucket = curr_bucket;
+#ifdef DEBUG
+					items_list = buckets[non_placed_bucket + sorted_lists[i].buckets_list].items_list;
+					bucket_id = buckets[non_placed_bucket + sorted_lists[i].buckets_list].bucket_id;
+#endif
+					buckets[non_placed_bucket + sorted_lists[i].buckets_list].items_list = buckets[curr_bucket].items_list;
+					buckets[non_placed_bucket + sorted_lists[i].buckets_list].bucket_id = buckets[curr_bucket].bucket_id;
+#ifdef DEBUG		
+					buckets[curr_bucket].items_list=items_list;
+					buckets[curr_bucket].bucket_id=bucket_id;
+#endif
+					non_placed_bucket++;
 				}
-				curr_bucket = buckets[curr_bucket].next_in_list;
+				curr_bucket++;
 			};
+			sorted_lists[i].size = non_placed_bucket;
 			probe0_num++;
 			if(probe0_num >= chd_ph->n)
 			{
@@ -432,58 +547,73 @@ static inline cmph_uint8 place_buckets2(chd_ph_config_data_t *chd_ph, chd_ph_buc
 			probe_num++;
 			if(probe_num >= max_probes || probe1_num >= chd_ph->n)
 			{
+				sorted_lists[i].size = sorted_list_size;
 				return 0;
 			};
 		};
+		sorted_lists[i].size = sorted_list_size;
 	};
 	return 1;
 };
 
-cmph_uint8 chd_ph_searching(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, cmph_uint32 max_bucket_size,
-                            cmph_uint32 *sorted_lists, cmph_uint32 max_probes, cmph_uint32 * disp_table)
+cmph_uint8 chd_ph_searching(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, chd_ph_item_t *items ,
+			    cmph_uint32 max_bucket_size, chd_ph_sorted_list_t *sorted_lists, cmph_uint32 max_probes, 
+			    cmph_uint32 * disp_table)
 {
 	if(chd_ph->use_h)
 	{
-		return place_buckets2(chd_ph, buckets, max_bucket_size, sorted_lists, max_probes, disp_table);
+		return place_buckets2(chd_ph, buckets, items, max_bucket_size, sorted_lists, max_probes, disp_table);
 	}
 	else
 	{
-		return place_buckets1(chd_ph, buckets, max_bucket_size, sorted_lists, max_probes, disp_table);
+		return place_buckets1(chd_ph, buckets, items, max_bucket_size, sorted_lists, max_probes, disp_table);
 	}
 
 }
 
-static inline cmph_uint8 chd_ph_check_bin_hashing(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, 
-                                                  cmph_uint32 * disp_table)
+static inline cmph_uint8 chd_ph_check_bin_hashing(chd_ph_config_data_t *chd_ph, chd_ph_bucket_t *buckets, chd_ph_item_t *items,
+                                                  cmph_uint32 * disp_table, chd_ph_sorted_list_t * sorted_lists,cmph_uint32 max_bucket_size)
 {
-	register cmph_uint32 i, j;
+	register cmph_uint32 bucket_size, i, j;
 	register cmph_uint32 position, probe0_num, probe1_num;
 	register cmph_uint32 m = 0;
 	register chd_ph_item_t * item;
-
-	memset(chd_ph->occup_table, 0, chd_ph->n);
-	for(i = 0; i < chd_ph->nbuckets; i++)
-	{
-		j = buckets[i].size;
-		item = buckets[i].items_list;
-		probe0_num = disp_table[i] % chd_ph->n;
-		probe1_num = disp_table[i] / chd_ph->n;
-		for(; j > 0; j--)
+	if(chd_ph->keys_per_bin > 1)
+		memset(chd_ph->occup_table, 0, chd_ph->n);
+	else
+		memset(chd_ph->occup_table, 0, ((chd_ph->n + 31)/32) * sizeof(cmph_uint32));
+		
+	for(bucket_size = 1; bucket_size <= max_bucket_size; bucket_size++)
+		for(i = sorted_lists[bucket_size].buckets_list; i < sorted_lists[bucket_size].size +
+				sorted_lists[bucket_size].buckets_list; i++)
 		{
-			if(item == 0)
+			j = bucket_size;
+			item = items + buckets[i].items_list;
+			probe0_num = disp_table[buckets[i].bucket_id] % chd_ph->n;
+			probe1_num = disp_table[buckets[i].bucket_id] / chd_ph->n;
+			for(; j > 0; j--)
 			{
-				return 0;
-			}
-			m++;
-			position = (item->f + ((cmph_uint64 )item->h) * probe0_num + probe1_num) % chd_ph->n;
-			if(chd_ph->occup_table[position] >= chd_ph->keys_per_bin)
-			{
-				return 0;
-			}
-			(chd_ph->occup_table[position])++;
-			item = item->next;
+				m++;
+				position = (item->f + ((cmph_uint64 )item->h) * probe0_num + probe1_num) % chd_ph->n;
+				if(chd_ph->keys_per_bin > 1)
+				{
+					if(chd_ph->occup_table[position] >= chd_ph->keys_per_bin)
+					{
+						return 0;
+					}
+					(chd_ph->occup_table[position])++;
+				} 
+				else
+				{
+					if(GETBIT32(((cmph_uint32*)chd_ph->occup_table), position))
+					{
+						return 0;
+					}
+					SETBIT32(((cmph_uint32*)chd_ph->occup_table), position);
+				};
+				item++;
+			};
 		};
-	};
 	DEBUGP("We were able to place m = %u keys\n", m);
 	return 1;
 };
@@ -503,7 +633,7 @@ cmph_t *chd_ph_new(cmph_config_t *mph, double c)
 	chd_ph_item_t * items = NULL;
 	register cmph_uint8 failure = 0;
 	cmph_uint32 max_bucket_size = 0;
-	cmph_uint32 * sorted_lists = NULL;
+	chd_ph_sorted_list_t * sorted_lists = NULL;
 	cmph_uint32 * disp_table = NULL;
 	register double space_lower_bound = 0;
 	#ifdef CMPH_TIMING
@@ -559,7 +689,12 @@ cmph_t *chd_ph_new(cmph_config_t *mph, double c)
 	items   = (chd_ph_item_t *) calloc(chd_ph->m, sizeof(chd_ph_item_t));
 
 	max_probes = (cmph_uint32)(((log(chd_ph->m)/log(2))/20) * max_probes);
-	chd_ph->occup_table = (cmph_uint8 *) calloc(chd_ph->n, sizeof(cmph_uint8));
+	
+	if(chd_ph->keys_per_bin == 1)
+		chd_ph->occup_table = (cmph_uint8 *) calloc(((chd_ph->n + 31)/32), sizeof(cmph_uint32));
+	else
+		chd_ph->occup_table = (cmph_uint8 *) calloc(chd_ph->n, sizeof(cmph_uint8));
+		
 	disp_table = (cmph_uint32 *) calloc(chd_ph->nbuckets, sizeof(cmph_uint32));
 // 	
 // 	init_genrand(time(0));
@@ -590,19 +725,22 @@ cmph_t *chd_ph_new(cmph_config_t *mph, double c)
 		{
 			free(sorted_lists);
 		}
-        	sorted_lists = chd_ph_ordering(buckets, chd_ph->nbuckets, max_bucket_size);
+
+        	sorted_lists = chd_ph_ordering(&buckets, &items, chd_ph->nbuckets, chd_ph->m, max_bucket_size);
 		
 		if (mph->verbosity)
 		{
 			fprintf(stderr, "Starting searching step\n");
 		}
 		
-		searching_success = chd_ph_searching(chd_ph, buckets, max_bucket_size, sorted_lists, max_probes, disp_table);
-			
+		searching_success = chd_ph_searching(chd_ph, buckets, items, max_bucket_size, sorted_lists, max_probes, disp_table);
 		if(searching_success) break;
 		
 		// reset occup_table
-		memset(chd_ph->occup_table, 0, chd_ph->n);
+		if(chd_ph->keys_per_bin > 1)
+			memset(chd_ph->occup_table, 0, chd_ph->n);
+		else
+			memset(chd_ph->occup_table, 0, ((chd_ph->n + 31)/32) * sizeof(cmph_uint32));
 		if(iterations == 0)
 		{
 			// Cleanup memory
@@ -617,11 +755,12 @@ cmph_t *chd_ph_new(cmph_config_t *mph, double c)
 
 	#ifdef DEBUG
 	{
-		if(!chd_ph_check_bin_hashing(chd_ph, buckets, disp_table))
+		if(!chd_ph_check_bin_hashing(chd_ph, buckets, items, disp_table,sorted_lists,max_bucket_size))
 		{
 		
 			DEBUGP("Error for bin packing generation");
-			return NULL;
+			failure = 1;
+			goto cleanup;
 		}
 	}
 	#endif