-/* Copyright (c) 2004, 2005, 2006, 2007, 2008, 2009, 2010. The SimGrid Team.
+/* Copyright (c) 2004-2015. The SimGrid Team.
* All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
* under the terms of the license (GNU LGPL) which comes with this package. */
#include "msg_private.h"
+#include "simix/smx_private.h"
#include "xbt/sysdep.h"
#include "xbt/log.h"
/** @addtogroup m_task_management
- *
- *
+ *
+ *
* Since most scheduling algorithms rely on a concept of task
* that can be either <em>computed</em> locally or
* <em>transferred</em> on another processor, it seems to be the
corresponding object.
* \param name a name for the object. It is for user-level information
and can be NULL.
- * \param compute_duration a value of the processing amount (in flop)
+ * \param flop_amount a value of the processing amount (in flop)
needed to process this new task. If 0, then it cannot be executed with
MSG_task_execute(). This value has to be >=0.
* \param message_size a value of the amount of data (in bytes) needed to
* \see msg_task_t
* \return The new corresponding object.
*/
-msg_task_t MSG_task_create(const char *name, double compute_duration,
+msg_task_t MSG_task_create(const char *name, double flop_amount,
double message_size, void *data)
{
msg_task_t task = xbt_new(s_msg_task_t, 1);
/* Simulator Data */
simdata->compute = NULL;
simdata->comm = NULL;
- simdata->message_size = message_size;
- simdata->computation_amount = compute_duration;
+ simdata->bytes_amount = message_size;
+ simdata->flops_amount = flop_amount;
simdata->sender = NULL;
simdata->receiver = NULL;
simdata->source = NULL;
simdata->priority = 1.0;
+ simdata->bound = 0;
+ simdata->affinity_mask_db = xbt_dict_new_homogeneous(NULL);
simdata->rate = -1.0;
simdata->isused = 0;
simdata->host_nb = 0;
simdata->host_list = NULL;
- simdata->comp_amount = NULL;
- simdata->comm_amount = NULL;
-#ifdef HAVE_TRACING
+ simdata->flops_parallel_amount = NULL;
+ simdata->bytes_parallel_amount = NULL;
TRACE_msg_task_create(task);
-#endif
return task;
}
and can be NULL.
* \param host_nb the number of hosts implied in the parallel task.
* \param host_list an array of \p host_nb msg_host_t.
- * \param computation_amount an array of \p host_nb
- doubles. computation_amount[i] is the total number of operations
- that have to be performed on host_list[i].
- * \param communication_amount an array of \p host_nb* \p host_nb doubles.
- * \param data a pointer to any data may want to attach to the new
- object. It is for user-level information and can be NULL. It can
- be retrieved with the function \ref MSG_task_get_data.
+ * \param flops_amount an array of \p host_nb doubles.
+ * flops_amount[i] is the total number of operations that have to be performed on host_list[i].
+ * \param bytes_amount an array of \p host_nb* \p host_nb doubles.
+ * \param data a pointer to any data may want to attach to the new object.
+ * It is for user-level information and can be NULL.
+ * It can be retrieved with the function \ref MSG_task_get_data.
* \see msg_task_t
* \return The new corresponding object.
*/
msg_task_t
MSG_parallel_task_create(const char *name, int host_nb,
const msg_host_t * host_list,
- double *computation_amount,
- double *communication_amount, void *data)
+ double *flops_amount,
+ double *bytes_amount, void *data)
{
msg_task_t task = MSG_task_create(name, 0, 0, data);
simdata_task_t simdata = task->simdata;
/* Simulator Data specific to parallel tasks */
simdata->host_nb = host_nb;
- simdata->host_list = xbt_new0(smx_host_t, host_nb);
- simdata->comp_amount = computation_amount;
- simdata->comm_amount = communication_amount;
+ simdata->host_list = xbt_new0(sg_host_t, host_nb);
+ simdata->flops_parallel_amount = flops_amount;
+ simdata->bytes_parallel_amount = bytes_amount;
for (i = 0; i < host_nb; i++)
- simdata->host_list[i] = host_list[i]->smx_host;
+ simdata->host_list[i] = host_list[i];
return task;
}
* \param name a name for the object. It is for user-level information
and can be NULL.
- * \param compute_duration a value of the processing amount (in flop)
+ * \param flops_amount a value of the processing amount (in flop)
needed to process this new task. If 0, then it cannot be executed with
MSG_gpu_task_execute(). This value has to be >=0.
* \see msg_gpu_task_t
* \return The new corresponding object.
*/
-msg_gpu_task_t MSG_gpu_task_create(const char *name, double compute_duration,
+msg_gpu_task_t MSG_gpu_task_create(const char *name, double flops_amount,
double dispatch_latency, double collect_latency)
{
msg_gpu_task_t task = xbt_new(s_msg_gpu_task_t, 1);
task->name = xbt_strdup(name);
/* Simulator Data */
- simdata->computation_amount = compute_duration;
+ simdata->flops_amount = flops_amount;
simdata->dispatch_latency = dispatch_latency;
simdata->collect_latency = collect_latency;
-#ifdef HAVE_TRACING
- //FIXME
- /* TRACE_msg_gpu_task_create(task); */
-#endif
+ /* TRACE_msg_gpu_task_create(task); FIXME*/
return task;
}
*/
msg_error_t MSG_task_destroy(msg_task_t task)
{
- smx_action_t action = NULL;
+ smx_synchro_t action = NULL;
xbt_assert((task != NULL), "Invalid parameter");
if (task->simdata->isused) {
/* the task is being sent or executed: cancel it first */
MSG_task_cancel(task);
}
-#ifdef HAVE_TRACING
TRACE_msg_task_destroy(task);
-#endif
xbt_free(task->name);
action = task->simdata->compute;
if (action)
- simcall_host_execution_destroy(action);
+ simcall_process_execution_destroy(action);
/* parallel tasks only */
xbt_free(task->simdata->host_list);
+ xbt_dict_free(&task->simdata->affinity_mask_db);
+
/* free main structures */
xbt_free(task->simdata);
xbt_free(task);
*/
msg_error_t MSG_task_cancel(msg_task_t task)
{
- xbt_assert((task != NULL), "Invalid parameter");
+ xbt_assert((task != NULL), "Cannot cancel a NULL task");
if (task->simdata->compute) {
- simcall_host_execution_cancel(task->simdata->compute);
+ simcall_process_execution_cancel(task->simdata->compute);
}
else if (task->simdata->comm) {
- simcall_comm_cancel(task->simdata->comm);
- task->simdata->isused = 0;
+ simdata_task_t simdata = task->simdata;
+ simcall_comm_cancel(simdata->comm);
+ if (msg_global->debug_multiple_use && simdata->isused!=0)
+ xbt_ex_free(*(xbt_ex_t*)simdata->isused);
+ simdata->isused = 0;
}
return MSG_OK;
}
/** \ingroup m_task_management
- * \brief Returns the computation amount needed to process a task #msg_task_t.
+ * \brief Returns the remaining amount of flops needed to execute a task #msg_task_t.
*
* Once a task has been processed, this amount is set to 0. If you want, you
- * can reset this value with #MSG_task_set_compute_duration before restarting the task.
+ * can reset this value with #MSG_task_set_flops_amount before restarting the task.
*/
-double MSG_task_get_compute_duration(msg_task_t task)
-{
- xbt_assert((task != NULL)
- && (task->simdata != NULL), "Invalid parameter");
+double MSG_task_get_flops_amount(msg_task_t task) {
- return task->simdata->computation_amount;
+ if (task->simdata->compute) {
+ return simcall_process_execution_get_remains(task->simdata->compute);
+ } else {
+ return task->simdata->flops_amount;
+ }
}
* \brief set the computation amount needed to process a task #msg_task_t.
*
* \warning If the computation is ongoing (already started and not finished),
- * it is not modified by this call. And the termination of the ongoing task with
- * set the computation_amount to zero, overriding any value set during the
- * execution.
+ * it is not modified by this call. Moreover, after its completion, the ongoing
+ * execution with set the flops_amount to zero, overriding any value set during
+ * the execution.
*/
-void MSG_task_set_compute_duration(msg_task_t task,
- double computation_amount)
+void MSG_task_set_flops_amount(msg_task_t task, double flops_amount)
{
- xbt_assert(task, "Invalid parameter");
- task->simdata->computation_amount = computation_amount;
-
+ task->simdata->flops_amount = flops_amount;
}
/** \ingroup m_task_management
- * \brief Returns the remaining computation amount of a task #msg_task_t.
+ * \brief set the amount data attached with a task #msg_task_t.
*
- * If the task is ongoing, this call retrieves the remaining amount of work.
- * If it is not ongoing, it returns the total amount of work that will be
- * executed when the task starts.
+ * \warning If the transfer is ongoing (already started and not finished),
+ * it is not modified by this call.
*/
-double MSG_task_get_remaining_computation(msg_task_t task)
-{
- xbt_assert((task != NULL)
- && (task->simdata != NULL), "Invalid parameter");
- if (task->simdata->compute) {
- return simcall_host_execution_get_remains(task->simdata->compute);
- } else {
- return task->simdata->computation_amount;
- }
+void MSG_task_set_bytes_amount(msg_task_t task, double data_size)
+{
+ task->simdata->bytes_amount = data_size;
}
+
/** \ingroup m_task_management
* \brief Returns the total amount received by a task #msg_task_t.
* If the communication does not exist it will return 0.
* \brief Returns the size of the data attached to a task #msg_task_t.
*
*/
-double MSG_task_get_data_size(msg_task_t task)
+double MSG_task_get_bytes_amount(msg_task_t task)
{
xbt_assert((task != NULL)
&& (task->simdata != NULL), "Invalid parameter");
- return task->simdata->message_size;
+ return task->simdata->bytes_amount;
}
/** \ingroup m_task_management
- * \brief Changes the priority of a computation task. This priority doesn't affect
+ * \brief Changes the priority of a computation task. This priority doesn't affect
* the transfer rate. A priority of 2 will make a task receive two times more
* cpu power than the other ones.
*
task->simdata->priority = 1 / priority;
if (task->simdata->compute)
- simcall_host_execution_set_priority(task->simdata->compute,
+ simcall_process_execution_set_priority(task->simdata->compute,
task->simdata->priority);
}
+
+
+/** \ingroup m_task_management
+ * \brief Changes the maximum CPU utilization of a computation task.
+ * Unit is flops/s.
+ *
+ * For VMs, there is a pitfall. Please see MSG_vm_set_bound().
+ */
+void MSG_task_set_bound(msg_task_t task, double bound)
+{
+ xbt_assert(task, "Invalid parameter");
+ xbt_assert(task->simdata, "Invalid parameter");
+
+ if (bound == 0)
+ XBT_INFO("bound == 0 means no capping (i.e., unlimited).");
+
+ task->simdata->bound = bound;
+ if (task->simdata->compute)
+ simcall_process_execution_set_bound(task->simdata->compute,
+ task->simdata->bound);
+}
+
+
+/** \ingroup m_task_management
+ * \brief Changes the CPU affinity of a computation task.
+ *
+ * When pinning the given task to the first CPU core of the given host, use
+ * 0x01 for the mask value. Each bit of the mask value corresponds to each CPU
+ * core. See taskset(1) on Linux.
+ *
+ * \param task a target task
+ * \param host the host having a multi-core CPU
+ * \param mask the bit mask of a new CPU affinity setting for the task
+ *
+ *
+ * Usage:
+ * 0. Define a host with multiple cores.
+ * \<host id="PM0" power="1E8" core="2"/\>
+ *
+ * 1. Pin a given task to the first CPU core of a host.
+ * MSG_task_set_affinity(task, pm0, 0x01);
+ *
+ * 2. Pin a given task to the third CPU core of a host. Turn on the third bit of the mask.
+ * MSG_task_set_affinity(task, pm0, 0x04); // 0x04 == 100B
+ *
+ * 3. Pin a given VM to the first CPU core of a host.
+ * MSG_vm_set_affinity(vm, pm0, 0x01);
+ *
+ * See examples/msg/cloud/multicore.c for more information.
+ *
+ *
+ * Note:
+ * 1. The current code does not allow an affinity of a task to multiple cores.
+ * The mask value 0x03 (i.e., a given task will be executed on the first core
+ * or the second core) is not allowed. The mask value 0x01 or 0x02 works. See
+ * cpu_cas01.c for details.
+ *
+ * 2. It is recommended to first compare simulation results in both the Lazy
+ * and Full calculation modes (using --cfg=cpu/optim:Full or not). Fix
+ * cpu_cas01.c if you find wrong results in the Lazy mode.
+ *
+ */
+void MSG_task_set_affinity(msg_task_t task, msg_host_t host, unsigned long mask)
+{
+ xbt_assert(task, "Invalid parameter");
+ xbt_assert(task->simdata, "Invalid parameter");
+
+ if (mask == 0) {
+ /* 0 means clear */
+ {
+ /* We need remove_ext() not throwing exception. */
+ void *ret = xbt_dict_get_or_null_ext(task->simdata->affinity_mask_db, (char *) host, sizeof(msg_host_t));
+ if (ret != NULL)
+ xbt_dict_remove_ext(task->simdata->affinity_mask_db, (char *) host, sizeof(host));
+ }
+ } else
+ xbt_dict_set_ext(task->simdata->affinity_mask_db, (char *) host, sizeof(host), (void *)(uintptr_t) mask, NULL);
+
+ /* We set affinity data of this task. If the task is being executed, we
+ * actually change the affinity setting of the task. Otherwise, this change
+ * will be applied when the task is executed. */
+
+ if (!task->simdata->compute) {
+ /* task is not yet executed */
+ XBT_INFO("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(host), MSG_task_get_name(task));
+ return;
+ }
+
+ {
+ smx_synchro_t compute = task->simdata->compute;
+ msg_host_t host_now = compute->execution.host; // simix_private.h is necessary
+ if (host_now != host) {
+ /* task is not yet executed on this host */
+ XBT_INFO("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(host), MSG_task_get_name(task));
+ return;
+ }
+
+ /* task is being executed on this host. so change the affinity now */
+ {
+ /* check it works. remove me if it works. */
+ xbt_assert((unsigned long)(uintptr_t) xbt_dict_get_or_null_ext(task->simdata->affinity_mask_db, (char *) host, sizeof(msg_host_t)) == mask);
+ }
+
+ XBT_INFO("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(host), MSG_task_get_name(task));
+ simcall_process_execution_set_affinity(task->simdata->compute, host, mask);
+ }
+}