diff --git a/include/haproxy/compiler.h b/include/haproxy/compiler.h
index f18b5fa5c..3943fe645 100644
--- a/include/haproxy/compiler.h
+++ b/include/haproxy/compiler.h
@@ -220,6 +220,17 @@
 #define HA_HAVE_CAS_DW
 #endif
 
+/*********************** IMPORTANT NOTE ABOUT ALIGNMENT **********************\
+ * Alignment works fine for variables. It also works on types and struct     *
+ * members by propagating the alignment to the container struct itself,      *
+ * but this requires that variables of the affected type are properly        *
+ * aligned themselves. While regular variables will always abide, those      *
+ * allocated using malloc() will not! Most platforms provide posix_memalign()*
+ * for this, but it's not available everywhere. As such one ought not to use *
+ * these alignment declarations inside structures that are dynamically       *
+ * allocated. If the purpose is only to avoid false sharing of cache lines   *
+ * for multi_threading, see THREAD_PAD() below.                              *
+\*****************************************************************************/
 
 /* sets alignment for current field or variable */
 #ifndef ALIGNED
@@ -294,6 +305,20 @@
 #endif
 #endif
 
+/* add optional padding of the specified size between fields in a structure,
+ * only when threads are enabled. This is used to avoid false sharing of cache
+ * lines for dynamically allocated structures which cannot guarantee alignment.
+ */
+#ifndef THREAD_PAD
+# ifdef USE_THREAD
+#  define __THREAD_PAD(x,l)  char __pad_##l[x]
+#  define _THREAD_PAD(x,l)   __THREAD_PAD(x, l)
+#  define THREAD_PAD(x)      _THREAD_PAD(x, __LINE__)
+# else
+#  define THREAD_PAD(x)
+# endif
+#endif
+
 /* The THREAD_LOCAL type attribute defines thread-local storage and is defined
  * to __thread when threads are enabled or empty when disabled.
  */
diff --git a/include/haproxy/server-t.h b/include/haproxy/server-t.h
index 09f39333d..e53d6121b 100644
--- a/include/haproxy/server-t.h
+++ b/include/haproxy/server-t.h
@@ -273,7 +273,7 @@ struct server {
 	/* The elements below may be changed on every single request by any
 	 * thread, and generally at the same time.
 	 */
-	ALWAYS_ALIGN(64);
+	THREAD_PAD(63);
 	struct eb32_node idle_node;             /* When to next do cleanup in the idle connections */
 	unsigned int curr_idle_conns;           /* Current number of orphan idling connections, both the idle and the safe lists */
 	unsigned int curr_idle_nb;              /* Current number of connections in the idle list */
@@ -288,14 +288,14 @@ struct server {
 	/* Element below are usd by LB algorithms and must be doable in
 	 * parallel to other threads reusing connections above.
 	 */
-	ALWAYS_ALIGN(64);
+	THREAD_PAD(63);
 	__decl_thread(HA_SPINLOCK_T lock);      /* may enclose the proxy's lock, must not be taken under */
 	unsigned npos, lpos;			/* next and last positions in the LB tree, protected by LB lock */
 	struct eb32_node lb_node;               /* node used for tree-based load balancing */
 	struct server *next_full;               /* next server in the temporary full list */
 
 	/* usually atomically updated by any thread during parsing or on end of request */
-	ALWAYS_ALIGN(64);
+	THREAD_PAD(63);
 	int cur_sess;				/* number of currently active sessions (including syn_sent) */
 	int served;				/* # of active sessions currently being served (ie not pending) */
 	int consecutive_errors;			/* current number of consecutive errors */
@@ -303,7 +303,7 @@ struct server {
 	struct be_counters counters;		/* statistics counters */
 
 	/* Below are some relatively stable settings, only changed under the lock */
-	ALWAYS_ALIGN(64);
+	THREAD_PAD(63);
 
 	struct eb_root *lb_tree;                /* we want to know in what tree the server is */
 	struct tree_occ *lb_nodes;              /* lb_nodes_tot * struct tree_occ */