Fix miscalculation of CPU shares on a multicore PM.

The recent commit adding the load capping feature broke CPU share
calculation on a multicore PM. This patch fixes the problem.

The bound value "0" means that there is no CPU bound for that task or
VM. On a single-core PM, this is true. The task can use all CPU resource
on that PM. On a multicore PM, the task can use the resource of
only a CPU core.

Even MSG_{task/vm}_set_bound() is not used in a simulation program,
simulation results will be wrong for multicore PMs. This is because
the bound value "0" is used internally.

diff --git a/src/msg/msg_vm.c b/src/msg/msg_vm.c
index 93afe69..bae0815 100644 (file)
--- a/src/msg/msg_vm.c
+++ b/src/msg/msg_vm.c
@@ -1250,7 +1250,9 @@ msg_host_t MSG_vm_get_pm(msg_vm_t vm)
  *
  *
  * 2.
- * Note that bound == 0 means no bound (i.e., unlimited).
+ * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
+ * multiple CPU cores, the CPU share of a computation task (or a VM) never
+ * exceeds the capacity of a CPU core.
  */
 void MSG_vm_set_bound(msg_vm_t vm, double bound)
 {

diff --git a/src/simix/smx_host.c b/src/simix/smx_host.c
index cf11967..177c4cd 100644 (file)
--- a/src/simix/smx_host.c
+++ b/src/simix/smx_host.c
@@ -423,7 +423,14 @@ smx_action_t SIMIX_host_execute(const char *name,
     action->execution.surf_exec = ws_model->extension.workstation.execute(host, computation_amount);
     ws_model->action_data_set(action->execution.surf_exec, action);
     ws_model->set_priority(action->execution.surf_exec, priority);
-    ws_model->set_bound(action->execution.surf_exec, bound);
+
+    /* Note (hypervisor): for multicore, the bound value being passed to the
+     * surf layer should not be zero (i.e., unlimited). It should be the
+     * capacity of a CPU core. */
+    if (bound == 0)
+      ws_model->set_bound(action->execution.surf_exec, SIMIX_host_get_speed(host));
+    else
+      ws_model->set_bound(action->execution.surf_exec, bound);
   }
 
   XBT_DEBUG("Create execute action %p", action);

diff --git a/src/surf/cpu_cas01.c b/src/surf/cpu_cas01.c
index cae096c..79fe3d6 100644 (file)
--- a/src/surf/cpu_cas01.c
+++ b/src/surf/cpu_cas01.c
@@ -237,6 +237,12 @@ static surf_action_t cpu_execute(void *cpu, double size)
   GENERIC_LMM_ACTION(action).suspended = 0;     /* Should be useless because of the
                                                    calloc but it seems to help valgrind... */
 
+  /* Note (hypervisor): here, the bound value of the variable is set to the
+   * capacity of a CPU core. But, after MSG_{task/vm}_set_bound() were added to
+   * the hypervisor branch, this bound value is overwritten in
+   * SIMIX_host_execute().
+   * TODO: cleanup this.
+   */
   GENERIC_LMM_ACTION(action).variable =
       lmm_variable_new(cpu_model->model_private->maxmin_system, action,
                        GENERIC_ACTION(action).priority,