Add some padding in the worker structure to make sure to separate out data in different cache lines

include/starpu_util.h

@@ -598,6 +598,17 @@ STARPU_ATOMIC_SOMETHING64(or, old | value)
 #define STARPU_WMB() STARPU_SYNCHRONIZE()
 #endif
 
+#if defined(__i386__) || defined(__x86_64__)
+#define STARPU_CACHELINE_SIZE 64
+#elif defined(__ppc__) || defined(__ppc64__) || defined(__ia64__)
+#define STARPU_CACHELINE_SIZE 128
+#elif defined(__s390__) || defined(__s390x__)
+#define STARPU_CACHELINE_SIZE 256
+#else
+/* Conservative default */
+#define STARPU_CACHELINE_SIZE 1024
+#endif
+
 #ifdef _WIN32
 /* Try to fetch the system definition of timespec */
 #include <sys/types.h>

src/core/workers.h

@@ -203,6 +203,10 @@ LIST_TYPE(_starpu_worker,
 
 	int enable_knob;
 	int bindid_requested;
+
+	/* Keep this last, to make sure to separate worker data in separate
+	  cache lines. */
+	char padding[STARPU_CACHELINE_SIZE];
 );
 
 struct _starpu_combined_worker
@@ -223,6 +227,10 @@ struct _starpu_combined_worker
 #ifdef STARPU_HAVE_HWLOC
 	hwloc_bitmap_t hwloc_cpu_set;
 #endif
+
+	/* Keep this last, to make sure to separate worker data in separate
+	  cache lines. */
+	char padding[STARPU_CACHELINE_SIZE];
 };
 
 /**

src/profiling/profiling.c

@@ -29,6 +29,8 @@
 #include <papi.h>
 #endif
 
+/* TODO: move to worker structure */
+
 static struct starpu_profiling_worker_info worker_info[STARPU_NMAXWORKERS];
 /* TODO: rather use rwlock */
 static starpu_pthread_mutex_t worker_info_mutex[STARPU_NMAXWORKERS];