bond: performance harvesting

- hash is great. But it is a bit too slow for the DP. Use direct array indexing
to quickly retrieve the slave interface.
- the algorithm used by flow hash is great. But it is a bit too slow for the DP.
Use l2_hash_hash() extracted from lb_hash.h which ECMP is using. It makes use
of intrinsic crc32 instruction set.
- shortcut modulo arithmetic when the operand is 2**x (where x up to 4) to
avoid division instruction.
- special case for link count == 1 in bond_tx_fn()
- use clib_mem_unaligned to access data for the packet to avoid alignment error
- Fix some typos for packet tracing.

Change-Id: I8eae3ad497061c5473aa675ba894ee0211120d25
Signed-off-by: Steven <sluong@cisco.com>

4 files changed, 183 insertions, 133 deletions

diff --git a/src/vnet/bonding/cli.c b/src/vnet/bonding/cli.c
index b66c4af337d..2799bb88b99 100644
--- a/src/vnet/bonding/cli.c
+++ b/src/vnet/bonding/cli.c
@@ -142,7 +142,7 @@ bond_delete_neighbor (vlib_main_t * vm, bond_if_t * bif, slave_if_t * sif)
   bif->port_number_bitmap =
     clib_bitmap_set (bif->port_number_bitmap,
 		     ntohs (sif->actor_admin.port_number) - 1, 0);
-  hash_unset (bm->neighbor_by_sw_if_index, sif->sw_if_index);
+  bm->slave_by_sw_if_index[sif->sw_if_index] = 0;
   vec_free (sif->last_marker_pkt);
   vec_free (sif->last_rx_pkt);
   vec_foreach_index (i, bif->slaves)
@@ -451,8 +451,15 @@ bond_enslave (vlib_main_t * vm, bond_enslave_args_t * args)
   else
     sif->ttl_in_seconds = LACP_SHORT_TIMOUT_TIME;
 
-  hash_set (bm->neighbor_by_sw_if_index, sif->sw_if_index,
-	    sif - bm->neighbors);
+  vec_validate_aligned (bm->slave_by_sw_if_index, sif->sw_if_index,
+			CLIB_CACHE_LINE_BYTES);
+  /*
+   * sif - bm->neighbors may be 0
+   * Left shift it by 1 bit to distinguish the valid entry that we actually
+   * store from the null entries
+   */
+  bm->slave_by_sw_if_index[sif->sw_if_index] =
+    (uword) (((sif - bm->neighbors) << 1) | 1);
   vec_add1 (bif->slaves, sif->sw_if_index);
 
   sif_hw = vnet_get_sup_hw_interface (vnm, sif->sw_if_index);
@@ -726,7 +733,7 @@ bond_cli_init (vlib_main_t * vm)
 
   bm->vlib_main = vm;
   bm->vnet_main = vnet_get_main ();
-  bm->neighbor_by_sw_if_index = hash_create (0, sizeof (uword));
+  vec_validate_aligned (bm->slave_by_sw_if_index, 1, CLIB_CACHE_LINE_BYTES);
 
   return 0;
 }
diff --git a/src/vnet/bonding/device.c b/src/vnet/bonding/device.c
index 9e1bb9517e4..8ddec80850a 100644
--- a/src/vnet/bonding/device.c
+++ b/src/vnet/bonding/device.c
@@ -22,6 +22,7 @@
 #include <vnet/ip/ip6_packet.h>
 #include <vnet/ip/ip6_hop_by_hop_packet.h>
 #include <vnet/bonding/node.h>
+#include <vppinfra/lb_hash_hash.h>
 
 #define foreach_bond_tx_error     \
   _(NONE, "no error")             \
@@ -126,7 +127,7 @@ bond_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, u32 flags)
   return 0;
 }
 
-static inline u32
+static_always_inline u32
 bond_load_balance_broadcast (vlib_main_t * vm, vlib_node_runtime_t * node,
 			     bond_if_t * bif, vlib_buffer_t * b0,
 			     uword slave_count)
@@ -160,29 +161,26 @@ bond_load_balance_broadcast (vlib_main_t * vm, vlib_node_runtime_t * node,
   return 0;
 }
 
-static inline u32
+static_always_inline u32
 bond_load_balance_l2 (vlib_main_t * vm, vlib_node_runtime_t * node,
 		      bond_if_t * bif, vlib_buffer_t * b0, uword slave_count)
 {
   ethernet_header_t *eth = (ethernet_header_t *) vlib_buffer_get_current (b0);
-  u32 a = 0, b = 0, c = 0, t1, t2;
-  u16 t11, t22;
+  u32 c;
+  u64 *dst = (u64 *) & eth->dst_address[0];
+  u64 a = clib_mem_unaligned (dst, u64);
+  u32 *src = (u32 *) & eth->src_address[2];
+  u32 b = clib_mem_unaligned (src, u32);
 
-  memcpy (&t1, eth->src_address, sizeof (t1));
-  memcpy (&t11, &eth->src_address[4], sizeof (t11));
-  a = t1 ^ t11;
+  c = lb_hash_hash_2_tuples (a, b);
 
-  memcpy (&t2, eth->dst_address, sizeof (t2));
-  memcpy (&t22, &eth->dst_address[4], sizeof (t22));
-  b = t2 ^ t22;
-
-  hash_v3_mix32 (a, b, c);
-  hash_v3_finalize32 (a, b, c);
-
-  return c % slave_count;
+  if (BOND_MODULO_SHORTCUT (slave_count))
+    return (c & (slave_count - 1));
+  else
+    return c % slave_count;
 }
 
-static inline u16 *
+static_always_inline u16 *
 bond_locate_ethertype (ethernet_header_t * eth)
 {
   u16 *ethertype_p;
@@ -205,7 +203,7 @@ bond_locate_ethertype (ethernet_header_t * eth)
   return ethertype_p;
 }
 
-static inline u32
+static_always_inline u32
 bond_load_balance_l23 (vlib_main_t * vm, vlib_node_runtime_t * node,
 		       bond_if_t * bif, vlib_buffer_t * b0, uword slave_count)
 {
@@ -213,9 +211,10 @@ bond_load_balance_l23 (vlib_main_t * vm, vlib_node_runtime_t * node,
   u8 ip_version;
   ip4_header_t *ip4;
   u16 ethertype, *ethertype_p;
+  u32 *mac1, *mac2, *mac3;
 
   ethertype_p = bond_locate_ethertype (eth);
-  ethertype = *ethertype_p;
+  ethertype = clib_mem_unaligned (ethertype_p, u16);
 
   if ((ethertype != htons (ETHERNET_TYPE_IP4)) &&
       (ethertype != htons (ETHERNET_TYPE_IP6)))
@@ -226,55 +225,63 @@ bond_load_balance_l23 (vlib_main_t * vm, vlib_node_runtime_t * node,
 
   if (ip_version == 0x4)
     {
-      u16 t11, t22;
-      u32 a = 0, b = 0, c = 0, t1, t2;
-
-      memcpy (&t1, eth->src_address, sizeof (t1));
-      memcpy (&t11, &eth->src_address[4], sizeof (t11));
-      a = t1 ^ t11;
-
-      memcpy (&t2, eth->dst_address, sizeof (t2));
-      memcpy (&t22, &eth->dst_address[4], sizeof (t22));
-      b = t2 ^ t22;
-
-      c = ip4->src_address.data_u32 ^ ip4->dst_address.data_u32;
-
-      hash_v3_mix32 (a, b, c);
-      hash_v3_finalize32 (a, b, c);
-
-      return c % slave_count;
+      u32 a, c;
+
+      mac1 = (u32 *) & eth->dst_address[0];
+      mac2 = (u32 *) & eth->dst_address[4];
+      mac3 = (u32 *) & eth->src_address[2];
+
+      a = clib_mem_unaligned (mac1, u32) ^ clib_mem_unaligned (mac2, u32) ^
+	clib_mem_unaligned (mac3, u32);
+      c =
+	lb_hash_hash_2_tuples (clib_mem_unaligned (&ip4->address_pair, u64),
+			       a);
+      if (BOND_MODULO_SHORTCUT (slave_count))
+	return (c & (slave_count - 1));
+      else
+	return c % slave_count;
     }
   else if (ip_version == 0x6)
     {
-      u64 a, b, c;
-      u64 t1 = 0, t2 = 0;
+      u64 a;
+      u32 c;
       ip6_header_t *ip6 = (ip6_header_t *) (eth + 1);
 
-      memcpy (&t1, eth->src_address, sizeof (eth->src_address));
-      memcpy (&t2, eth->dst_address, sizeof (eth->dst_address));
-      a = t1 ^ t2;
-
-      b = (ip6->src_address.as_u64[0] ^ ip6->src_address.as_u64[1]);
-      c = (ip6->dst_address.as_u64[0] ^ ip6->dst_address.as_u64[1]);
-
-      hash_mix64 (a, b, c);
-      return c % slave_count;
+      mac1 = (u32 *) & eth->dst_address[0];
+      mac2 = (u32 *) & eth->dst_address[4];
+      mac3 = (u32 *) & eth->src_address[2];
+
+      a = clib_mem_unaligned (mac1, u32) ^ clib_mem_unaligned (mac2, u32) ^
+	clib_mem_unaligned (mac3, u32);
+      c =
+	lb_hash_hash (clib_mem_unaligned
+		      (&ip6->src_address.as_uword[0], uword),
+		      clib_mem_unaligned (&ip6->src_address.as_uword[1],
+					  uword),
+		      clib_mem_unaligned (&ip6->dst_address.as_uword[0],
+					  uword),
+		      clib_mem_unaligned (&ip6->dst_address.as_uword[1],
+					  uword), a);
+      if (BOND_MODULO_SHORTCUT (slave_count))
+	return (c & (slave_count - 1));
+      else
+	return c % slave_count;
     }
   return (bond_load_balance_l2 (vm, node, bif, b0, slave_count));
 }
 
-static inline u32
+static_always_inline u32
 bond_load_balance_l34 (vlib_main_t * vm, vlib_node_runtime_t * node,
 		       bond_if_t * bif, vlib_buffer_t * b0, uword slave_count)
 {
   ethernet_header_t *eth = (ethernet_header_t *) vlib_buffer_get_current (b0);
   u8 ip_version;
-  uword is_tcp_udp = 0;
+  uword is_tcp_udp;
   ip4_header_t *ip4;
   u16 ethertype, *ethertype_p;
 
   ethertype_p = bond_locate_ethertype (eth);
-  ethertype = *ethertype_p;
+  ethertype = clib_mem_unaligned (ethertype_p, u16);
 
   if ((ethertype != htons (ETHERNET_TYPE_IP4)) &&
       (ethertype != htons (ETHERNET_TYPE_IP6)))
@@ -285,29 +292,30 @@ bond_load_balance_l34 (vlib_main_t * vm, vlib_node_runtime_t * node,
 
   if (ip_version == 0x4)
     {
-      u32 a = 0, b = 0, c = 0, t1, t2;
+      u32 a, c, t1, t2;
       tcp_header_t *tcp = (void *) (ip4 + 1);
+
       is_tcp_udp = (ip4->protocol == IP_PROTOCOL_TCP) ||
 	(ip4->protocol == IP_PROTOCOL_UDP);
-
-      a = ip4->src_address.data_u32 ^ ip4->dst_address.data_u32;
-
-      t1 = is_tcp_udp ? tcp->src : 0;
-      t2 = is_tcp_udp ? tcp->dst : 0;
-      b = t1 + (t2 << 16);
-
-      hash_v3_mix32 (a, b, c);
-      hash_v3_finalize32 (a, b, c);
-
-      return c % slave_count;
+      t1 = is_tcp_udp ? clib_mem_unaligned (&tcp->src, u16) : 0;
+      t2 = is_tcp_udp ? clib_mem_unaligned (&tcp->dst, u16) : 0;
+      a = t1 ^ t2;
+      c =
+	lb_hash_hash_2_tuples (clib_mem_unaligned (&ip4->address_pair, u64),
+			       a);
+      if (BOND_MODULO_SHORTCUT (slave_count))
+	return (c & (slave_count - 1));
+      else
+	return c % slave_count;
     }
   else if (ip_version == 0x6)
     {
-      u64 a, b, c;
-      u64 t1, t2;
+      u64 a;
+      u32 c, t1, t2;
       ip6_header_t *ip6 = (ip6_header_t *) (eth + 1);
       tcp_header_t *tcp = (void *) (ip6 + 1);
 
+      is_tcp_udp = 0;
       if (PREDICT_TRUE ((ip6->protocol == IP_PROTOCOL_TCP) ||
 			(ip6->protocol == IP_PROTOCOL_UDP)))
 	{
@@ -325,33 +333,43 @@ bond_load_balance_l34 (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      tcp = (tcp_header_t *) ((u8 *) hbh + ((hbh->length + 1) << 3));
 	    }
 	}
-      a = (ip6->src_address.as_u64[0] ^ ip6->src_address.as_u64[1]);
-      b = (ip6->dst_address.as_u64[0] ^ ip6->dst_address.as_u64[1]);
-
-      t1 = is_tcp_udp ? tcp->src : 0;
-      t2 = is_tcp_udp ? tcp->dst : 0;
-      c = (t2 << 16) | t1;
-      hash_mix64 (a, b, c);
-
-      return c % slave_count;
+      t1 = is_tcp_udp ? clib_mem_unaligned (&tcp->src, u16) : 0;
+      t2 = is_tcp_udp ? clib_mem_unaligned (&tcp->dst, u16) : 0;
+      a = t1 ^ t2;
+      c =
+	lb_hash_hash (clib_mem_unaligned
+		      (&ip6->src_address.as_uword[0], uword),
+		      clib_mem_unaligned (&ip6->src_address.as_uword[1],
+					  uword),
+		      clib_mem_unaligned (&ip6->dst_address.as_uword[0],
+					  uword),
+		      clib_mem_unaligned (&ip6->dst_address.as_uword[1],
+					  uword), a);
+      if (BOND_MODULO_SHORTCUT (slave_count))
+	return (c & (slave_count - 1));
+      else
+	return c % slave_count;
     }
 
   return (bond_load_balance_l2 (vm, node, bif, b0, slave_count));
 }
 
-static inline u32
+static_always_inline u32
 bond_load_balance_round_robin (vlib_main_t * vm,
 			       vlib_node_runtime_t * node,
 			       bond_if_t * bif, vlib_buffer_t * b0,
 			       uword slave_count)
 {
   bif->lb_rr_last_index++;
-  bif->lb_rr_last_index %= slave_count;
+  if (BOND_MODULO_SHORTCUT (slave_count))
+    bif->lb_rr_last_index &= slave_count - 1;
+  else
+    bif->lb_rr_last_index %= slave_count;
 
   return bif->lb_rr_last_index;
 }
 
-static inline u32
+static_always_inline u32
 bond_load_balance_active_backup (vlib_main_t * vm,
 				 vlib_node_runtime_t * node,
 				 bond_if_t * bif, vlib_buffer_t * b0,
@@ -379,8 +397,7 @@ bond_tx_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
   u32 *from = vlib_frame_vector_args (frame);
   u32 n_left_from;
   ethernet_header_t *eth;
-  u32 next0 = 0, next1 = 0, next2 = 0, next3 = 0;
-  u32 port, port1, port2, port3;
+  u32 port;
   u32 sw_if_index, sw_if_index1, sw_if_index2, sw_if_index3;
   bond_packet_trace_t *t0;
   uword n_trace = vlib_get_trace_count (vm, node);
@@ -435,6 +452,9 @@ bond_tx_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
     {
       while (n_left_from >= 4)
 	{
+	  u32 next0 = 0, next1 = 0, next2 = 0, next3 = 0;
+	  u32 port0 = 0, port1 = 0, port2 = 0, port3 = 0;
+
 	  // Prefetch next iteration
 	  if (n_left_from >= 8)
 	    {
@@ -445,10 +465,10 @@ bond_tx_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      p6 = vlib_get_buffer (vm, from[6]);
 	      p7 = vlib_get_buffer (vm, from[7]);
 
-	      vlib_prefetch_buffer_header (p4, STORE);
-	      vlib_prefetch_buffer_header (p5, STORE);
-	      vlib_prefetch_buffer_header (p6, STORE);
-	      vlib_prefetch_buffer_header (p7, STORE);
+	      vlib_prefetch_buffer_header (p4, LOAD);
+	      vlib_prefetch_buffer_header (p5, LOAD);
+	      vlib_prefetch_buffer_header (p6, LOAD);
+	      vlib_prefetch_buffer_header (p7, LOAD);
 
 	      CLIB_PREFETCH (p4->data, CLIB_CACHE_LINE_BYTES, LOAD);
 	      CLIB_PREFETCH (p5->data, CLIB_CACHE_LINE_BYTES, LOAD);
@@ -476,20 +496,27 @@ bond_tx_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  sw_if_index2 = vnet_buffer (b2)->sw_if_index[VLIB_TX];
 	  sw_if_index3 = vnet_buffer (b3)->sw_if_index[VLIB_TX];
 
-	  port =
-	    (bond_load_balance_table[bif->lb]).load_balance (vm, node, bif,
-							     b0, slave_count);
-	  port1 =
-	    (bond_load_balance_table[bif->lb]).load_balance (vm, node, bif,
-							     b1, slave_count);
-	  port2 =
-	    (bond_load_balance_table[bif->lb]).load_balance (vm, node, bif,
-							     b2, slave_count);
-	  port3 =
-	    (bond_load_balance_table[bif->lb]).load_balance (vm, node, bif,
-							     b3, slave_count);
-
-	  sif_if_index = *vec_elt_at_index (bif->active_slaves, port);
+	  if (PREDICT_TRUE (slave_count != 1))
+	    {
+	      port0 =
+		(bond_load_balance_table[bif->lb]).load_balance (vm, node,
+								 bif, b0,
+								 slave_count);
+	      port1 =
+		(bond_load_balance_table[bif->lb]).load_balance (vm, node,
+								 bif, b1,
+								 slave_count);
+	      port2 =
+		(bond_load_balance_table[bif->lb]).load_balance (vm, node,
+								 bif, b2,
+								 slave_count);
+	      port3 =
+		(bond_load_balance_table[bif->lb]).load_balance (vm, node,
+								 bif, b3,
+								 slave_count);
+	    }
+
+	  sif_if_index = *vec_elt_at_index (bif->active_slaves, port0);
 	  sif_if_index1 = *vec_elt_at_index (bif->active_slaves, port1);
 	  sif_if_index2 = *vec_elt_at_index (bif->active_slaves, port2);
 	  sif_if_index3 = *vec_elt_at_index (bif->active_slaves, port3);
@@ -499,23 +526,27 @@ bond_tx_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  vnet_buffer (b2)->sw_if_index[VLIB_TX] = sif_if_index2;
 	  vnet_buffer (b3)->sw_if_index[VLIB_TX] = sif_if_index3;
 
-	  if (bif->per_thread_info[thread_index].frame[port] == 0)
-	    bif->per_thread_info[thread_index].frame[port] =
+	  if (PREDICT_FALSE ((bif->per_thread_info[thread_index].frame[port0]
+			      == 0)))
+	    bif->per_thread_info[thread_index].frame[port0] =
 	      vnet_get_frame_to_sw_interface (vnm, sif_if_index);
 
-	  if (bif->per_thread_info[thread_index].frame[port1] == 0)
+	  if (PREDICT_FALSE ((bif->per_thread_info[thread_index].frame[port1]
+			      == 0)))
 	    bif->per_thread_info[thread_index].frame[port1] =
 	      vnet_get_frame_to_sw_interface (vnm, sif_if_index1);
 
-	  if (bif->per_thread_info[thread_index].frame[port2] == 0)
+	  if (PREDICT_FALSE ((bif->per_thread_info[thread_index].frame[port2]
+			      == 0)))
 	    bif->per_thread_info[thread_index].frame[port2] =
 	      vnet_get_frame_to_sw_interface (vnm, sif_if_index2);
 
-	  if (bif->per_thread_info[thread_index].frame[port3] == 0)
+	  if (PREDICT_FALSE ((bif->per_thread_info[thread_index].frame[port3]
+			      == 0)))
 	    bif->per_thread_info[thread_index].frame[port3] =
 	      vnet_get_frame_to_sw_interface (vnm, sif_if_index3);
 
-	  f = bif->per_thread_info[thread_index].frame[port];
+	  f = bif->per_thread_info[thread_index].frame[port0];
 	  to_next = vlib_frame_vector_args (f);
 	  to_next += f->n_vectors;
 	  to_next[0] = vlib_get_buffer_index (vm, b0);
@@ -597,13 +628,16 @@ bond_tx_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 
       while (n_left_from > 0)
 	{
+	  u32 next0 = 0;
+	  u32 port0 = 0;
+
 	  // Prefetch next iteration
 	  if (n_left_from > 1)
 	    {
 	      vlib_buffer_t *p2;
 
 	      p2 = vlib_get_buffer (vm, from[1]);
-	      vlib_prefetch_buffer_header (p2, STORE);
+	      vlib_prefetch_buffer_header (p2, LOAD);
 	      CLIB_PREFETCH (p2->data, CLIB_CACHE_LINE_BYTES, LOAD);
 	    }
 
@@ -614,15 +648,18 @@ bond_tx_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  sw_if_index = vnet_buffer (b0)->sw_if_index[VLIB_TX];
 
-	  port =
-	    (bond_load_balance_table[bif->lb]).load_balance (vm, node, bif,
-							     b0, slave_count);
-	  sif_if_index = *vec_elt_at_index (bif->active_slaves, port);
+	  if (PREDICT_TRUE (slave_count != 1))
+	    port0 =
+	      (bond_load_balance_table[bif->lb]).load_balance (vm, node, bif,
+							       b0,
+							       slave_count);
+	  sif_if_index = *vec_elt_at_index (bif->active_slaves, port0);
 	  vnet_buffer (b0)->sw_if_index[VLIB_TX] = sif_if_index;
-	  if (bif->per_thread_info[thread_index].frame[port] == 0)
-	    bif->per_thread_info[thread_index].frame[port] =
+	  if (PREDICT_FALSE
+	      ((bif->per_thread_info[thread_index].frame[port0] == 0)))
+	    bif->per_thread_info[thread_index].frame[port0] =
 	      vnet_get_frame_to_sw_interface (vnm, sif_if_index);
-	  f = bif->per_thread_info[thread_index].frame[port];
+	  f = bif->per_thread_info[thread_index].frame[port0];
 	  to_next = vlib_frame_vector_args (f);
 	  to_next += f->n_vectors;
 	  to_next[0] = vlib_get_buffer_index (vm, b0);
diff --git a/src/vnet/bonding/node.c b/src/vnet/bonding/node.c
index 65d3ba10470..ec251f550cf 100644
--- a/src/vnet/bonding/node.c
+++ b/src/vnet/bonding/node.c
@@ -90,20 +90,21 @@ bond_sw_if_index_rewrite (vlib_main_t * vm, vlib_node_runtime_t * node,
 
   if (PREDICT_TRUE (sif != 0))
     {
-      bif = bond_get_master_by_sw_if_index (sif->group);
+      bif = bond_get_master_by_dev_instance (sif->bif_dev_instance);
       if (PREDICT_TRUE (bif != 0))
 	{
 	  if (PREDICT_TRUE (vec_len (bif->slaves) >= 1))
 	    {
 	      if (PREDICT_TRUE (bif->admin_up == 1))
 		{
-		  if (!ethernet_frame_is_tagged (ntohs (eth->type)))
+		  ethertype = clib_mem_unaligned (&eth->type, u16);
+		  if (!ethernet_frame_is_tagged (ntohs (ethertype)))
 		    {
 		      // Let some layer2 packets pass through.
-		      if (PREDICT_TRUE ((eth->type !=
+		      if (PREDICT_TRUE ((ethertype !=
 					 htons (ETHERNET_TYPE_SLOW_PROTOCOLS))
 					&& !packet_is_cdp (eth)
-					&& (eth->type !=
+					&& (ethertype !=
 					    htons
 					    (ETHERNET_TYPE_802_1_LLDP))))
 			{
@@ -128,12 +129,13 @@ bond_sw_if_index_rewrite (vlib_main_t * vm, vlib_node_runtime_t * node,
 		    {
 		      vlan = (void *) (eth + 1);
 		      ethertype_p = &vlan->type;
-		      if (*ethertype_p == ntohs (ETHERNET_TYPE_VLAN))
+		      ethertype = clib_mem_unaligned (ethertype_p, u16);
+		      if (ethertype == ntohs (ETHERNET_TYPE_VLAN))
 			{
 			  vlan++;
 			  ethertype_p = &vlan->type;
 			}
-		      ethertype = *ethertype_p;
+		      ethertype = clib_mem_unaligned (ethertype_p, u16);
 		      if (PREDICT_TRUE ((ethertype !=
 					 htons (ETHERNET_TYPE_SLOW_PROTOCOLS))
 					&& (ethertype !=
@@ -226,10 +228,10 @@ bond_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    b6 = vlib_get_buffer (vm, from[6]);
 	    b7 = vlib_get_buffer (vm, from[7]);
 
-	    vlib_prefetch_buffer_header (b4, STORE);
-	    vlib_prefetch_buffer_header (b5, STORE);
-	    vlib_prefetch_buffer_header (b6, STORE);
-	    vlib_prefetch_buffer_header (b7, STORE);
+	    vlib_prefetch_buffer_header (b4, LOAD);
+	    vlib_prefetch_buffer_header (b5, LOAD);
+	    vlib_prefetch_buffer_header (b6, LOAD);
+	    vlib_prefetch_buffer_header (b7, LOAD);
 
 	    CLIB_PREFETCH (b4->data, CLIB_CACHE_LINE_BYTES, LOAD);
 	    CLIB_PREFETCH (b5->data, CLIB_CACHE_LINE_BYTES, LOAD);
@@ -314,7 +316,7 @@ bond_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 		  if (PREDICT_TRUE (n_trace > 0))
 		    {
-		      vlib_trace_buffer (vm, node, next1, b2,
+		      vlib_trace_buffer (vm, node, next2, b2,
 					 0 /* follow_chain */ );
 		      vlib_set_trace_count (vm, node, --n_trace);
 		      t0 = vlib_add_trace (vm, node, b2, sizeof (*t0));
@@ -325,7 +327,7 @@ bond_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 		      if (PREDICT_TRUE (n_trace > 0))
 			{
-			  vlib_trace_buffer (vm, node, next1, b2,
+			  vlib_trace_buffer (vm, node, next3, b3,
 					     0 /* follow_chain */ );
 			  vlib_set_trace_count (vm, node, --n_trace);
 			  t0 = vlib_add_trace (vm, node, b3, sizeof (*t0));
@@ -358,7 +360,7 @@ bond_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      vlib_buffer_t *p2;
 
 	      p2 = vlib_get_buffer (vm, from[1]);
-	      vlib_prefetch_buffer_header (p2, STORE);
+	      vlib_prefetch_buffer_header (p2, LOAD);
 	      CLIB_PREFETCH (p2->data, CLIB_CACHE_LINE_BYTES, LOAD);
 	    }
 
diff --git a/src/vnet/bonding/node.h b/src/vnet/bonding/node.h
index 3a01abe2226..5c6ff32ac4a 100644
--- a/src/vnet/bonding/node.h
+++ b/src/vnet/bonding/node.h
@@ -31,6 +31,9 @@
 #define MIN(x,y) (((x)<(y))?(x):(y))
 #endif
 
+#define BOND_MODULO_SHORTCUT(a) \
+  (((a) == 2) || ((a) == 4) || ((a) == 8) || ((a) == 16))
+
 #define foreach_bond_mode	    \
   _ (1, ROUND_ROBIN, "round-robin") \
   _ (2, ACTIVE_BACKUP, "active-backup") \
@@ -289,9 +292,6 @@ typedef struct
   /* pool of lacp neighbors */
   slave_if_t *neighbors;
 
-  /* rapidly find a neighbor by vlib software interface index */
-  uword *neighbor_by_sw_if_index;
-
   /* rapidly find a bond by vlib software interface index */
   uword *bond_by_sw_if_index;
 
@@ -303,6 +303,8 @@ typedef struct
   u8 lacp_plugin_loaded;
 
   lacp_enable_disable_func lacp_enable_disable;
+
+  uword *slave_by_sw_if_index;
 } bond_main_t;
 
 /* bond packet trace capture */
@@ -439,13 +441,15 @@ bond_get_slave_by_sw_if_index (u32 sw_if_index)
 {
   bond_main_t *bm = &bond_main;
   slave_if_t *sif = 0;
-  uword *p;
+  uword p;
 
-  p = hash_get (bm->neighbor_by_sw_if_index, sw_if_index);
-  if (p)
+  if (sw_if_index < vec_len (bm->slave_by_sw_if_index))
     {
-      sif = pool_elt_at_index (bm->neighbors, p[0]);
+      p = bm->slave_by_sw_if_index[sw_if_index];
+      if (p)
+	sif = pool_elt_at_index (bm->neighbors, p >> 1);
     }
+
   return sif;
 }