*
*/
msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name,
- int ncpus, int ramsize, int net_cap, char *disk_path, int disksize)
+ int ncpus, long ramsize, long net_cap, char *disk_path, long disksize)
{
msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
}
+static void launch_deferred_exec_process(msg_host_t host, double computation, double prio);
+
static int migration_rx_fun(int argc, char *argv[])
{
- const char *pr_name = MSG_process_get_name(MSG_process_self());
- const char *host_name = MSG_host_get_name(MSG_host_self());
-
XBT_DEBUG("mig: rx_start");
xbt_assert(argc == 4);
msg_vm_t vm = MSG_get_host_by_name(vm_name);
msg_vm_t dst_pm = MSG_get_host_by_name(dst_pm_name);
+
+ s_ws_params_t params;
+ simcall_host_get_params(vm, ¶ms);
+ const double xfer_cpu_overhead = params.xfer_cpu_overhead;
+
+
int need_exit = 0;
char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
for (;;) {
msg_task_t task = NULL;
MSG_task_recv(&task, mbox);
+ {
+ double received = MSG_task_get_data_size(task);
+ /* TODO: clean up */
+ // const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
+ launch_deferred_exec_process(vm, received * xfer_cpu_overhead, 1);
+ }
if (strcmp(task->name, finalize_task_name) == 0)
need_exit = 1;
return 0;
}
-
-typedef struct dirty_page {
- double prev_clock;
- double prev_remaining;
- msg_task_t task;
-} s_dirty_page, *dirty_page_t;
-
-
static void reset_dirty_pages(msg_vm_t vm)
{
msg_host_priv_t priv = msg_host_resource_priv(vm);
return computed;
}
-static double lookup_computed_flop_counts(msg_vm_t vm, int stage2_round_for_fancy_debug)
+static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
{
msg_host_priv_t priv = msg_host_resource_priv(vm);
double total = 0;
total += priv->dp_updated_by_deleted_tasks;
- XBT_INFO("mig-stage2.%d: computed %f flop_counts (including %f by deleted tasks)",
+ XBT_INFO("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
+ stage_for_fancy_debug,
stage2_round_for_fancy_debug,
total, priv->dp_updated_by_deleted_tasks);
}
+static int deferred_exec_fun(int argc, char *argv[])
+{
+ xbt_assert(argc == 3);
+ const char *comp_str = argv[1];
+ double computaion = atof(comp_str);
+ const char *prio_str = argv[2];
+ double prio = atof(prio_str);
+
+ msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL);
+ // XBT_INFO("exec deferred %f", computaion);
+
+ /* dpt is the results of the VM activity */
+ MSG_task_set_priority(task, prio);
+ MSG_task_execute(task);
+
+
+
+ MSG_task_destroy(task);
+
+ return 0;
+}
+
+static void launch_deferred_exec_process(msg_host_t host, double computation, double prio)
+{
+ char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host));
+
+ int nargvs = 4;
+ char **argv = xbt_new(char *, nargvs);
+ argv[0] = xbt_strdup(pr_name);
+ argv[1] = bprintf("%lf", computation);
+ argv[2] = bprintf("%lf", prio);
+ argv[3] = NULL;
+
+ msg_process_t pr = MSG_process_create_with_arguments(pr_name, deferred_exec_fun, NULL, host, nargvs - 1, argv);
+
+ xbt_free(pr_name);
+}
+
+
+static int task_tx_overhead_fun(int argc, char *argv[])
+{
+ xbt_assert(argc == 2);
+ const char *mbox = argv[1];
+
+ int need_exit = 0;
+
+ // XBT_INFO("start %s", mbox);
+
+ for (;;) {
+ msg_task_t task = NULL;
+ MSG_task_recv(&task, mbox);
+
+ // XBT_INFO("task->name %s", task->name);
+
+ if (strcmp(task->name, "finalize_making_overhead") == 0)
+ need_exit = 1;
+
+ // XBT_INFO("exec");
+ // MSG_task_set_priority(task, 1000000);
+ MSG_task_execute(task);
+ MSG_task_destroy(task);
+
+ if (need_exit)
+ break;
+ }
+
+ // XBT_INFO("bye");
+
+ return 0;
+}
+
+static void start_overhead_process(msg_task_t comm_task)
+{
+ char *pr_name = bprintf("__pr_task_tx_overhead_%s", MSG_task_get_name(comm_task));
+ char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
+
+ int nargvs = 3;
+ char **argv = xbt_new(char *, nargvs);
+ argv[0] = xbt_strdup(pr_name);
+ argv[1] = xbt_strdup(mbox);
+ argv[2] = NULL;
+
+ // XBT_INFO("micro start: mbox %s", mbox);
+ msg_process_t pr = MSG_process_create_with_arguments(pr_name, task_tx_overhead_fun, NULL, MSG_host_self(), nargvs - 1, argv);
+
+ xbt_free(pr_name);
+ xbt_free(mbox);
+}
+
+static void shutdown_overhead_process(msg_task_t comm_task)
+{
+ char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
+
+ msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL);
+
+ // XBT_INFO("micro shutdown: mbox %s", mbox);
+ msg_error_t ret = MSG_task_send(task, mbox);
+ xbt_assert(ret == MSG_OK);
+
+ xbt_free(mbox);
+ // XBT_INFO("shutdown done");
+}
+
+static void request_overhead(msg_task_t comm_task, double computation)
+{
+ char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
+
+ msg_task_t task = MSG_task_create("micro", computation, 0, NULL);
+
+ // XBT_INFO("req overhead");
+ msg_error_t ret = MSG_task_send(task, mbox);
+ xbt_assert(ret == MSG_OK);
+
+ xbt_free(mbox);
+}
+
+/* alpha is (floating_operations / bytes).
+ *
+ * When actual migration traffic was 32 mbytes/s, we observed the CPU
+ * utilization of the main thread of the Qemu process was 10 %.
+ * alpha = 0.1 * C / (32 * 1024 * 1024)
+ * where the CPU capacity of the PM is C flops/s.
+ *
+ * */
+static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha)
+{
+ const double chunk_size = 1024 * 1024 * 10;
+ double remaining = MSG_task_get_data_size(comm_task);
+
+ start_overhead_process(comm_task);
+
+
+ while (remaining > 0) {
+ double data_size = chunk_size;
+ if (remaining < chunk_size)
+ data_size = remaining;
+
+ remaining -= data_size;
+
+ // XBT_INFO("remaining %f bytes", remaining);
+
+
+ double clock_sta = MSG_get_clock();
+
+ /* create a micro task */
+ {
+ char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
+ msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL);
+
+ request_overhead(comm_task, data_size * alpha);
+
+ msg_error_t ret = MSG_task_send(mtask, mbox);
+ xbt_assert(ret == MSG_OK);
+
+ xbt_free(mtask_name);
+ }
+
+#if 0
+ {
+ /* In the real world, sending data involves small CPU computation. */
+ char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
+ msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL);
+ MSG_task_execute(mtask);
+ MSG_task_destroy(mtask);
+ xbt_free(mtask_name);
+ }
+#endif
+
+ /* TODO */
+
+ double clock_end = MSG_get_clock();
+
+
+ if (mig_speed > 0) {
+ /*
+ * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep)
+ *
+ * Thus, we get
+ * time_to_sleep > data_size / (max bandwidth) - (elapsed time)
+ *
+ * If time_to_sleep is smaller than zero, the elapsed time was too big. We
+ * do not need a micro sleep.
+ **/
+ double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta);
+ if (time_to_sleep > 0)
+ MSG_process_sleep(time_to_sleep);
+
+
+ //XBT_INFO("duration %f", clock_end - clock_sta);
+ //XBT_INFO("time_to_sleep %f", time_to_sleep);
+ }
+ }
+
+ // XBT_INFO("%s", MSG_task_get_name(comm_task));
+ shutdown_overhead_process(comm_task);
+
+}
+
+
+#if 0
+static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size)
+{
+ double prev_remaining = init_comm_size;
+
+ for (;;) {
+ double remaining = MSG_task_get_remaining_communication(comm_task);
+ if (remaining == 0)
+ need_exit = 1;
+
+ double sent = prev_remaining - remaining;
+ double comp_size = sent * overhead;
+
+
+ char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task));
+ msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL);
+ MSG_task_execute(comp_task);
+ MSG_task_destroy(comp_task);
+
+ if (need_exit)
+ break;
+
+ prev_remaining = remaining;
+
+ }
+
+ xbt_free(comp_task_name);
+}
+#endif
+
+#define USE_MICRO_TASK 1
+
+#if 0
+// const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024);
+// const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
+// const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024);
+// const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
+// const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s
+//const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024);
+// const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024);
+////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
+// const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024);
+// const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024);
+// const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
+
+/* CPU 22% when 80Mbyte/s */
+const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
+#endif
+
+
static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
- double size, char *mbox, int stage, int stage2_round, double mig_speed)
+ double size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead)
{
char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
+ /* TODO: clean up */
+
+ double clock_sta = MSG_get_clock();
+
+#ifdef USE_MICRO_TASK
+
+ task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead);
+
+#else
msg_error_t ret;
if (mig_speed > 0)
ret = MSG_task_send_bounded(task, mbox, mig_speed);
else
ret = MSG_task_send(task, mbox);
xbt_assert(ret == MSG_OK);
+#endif
+
+ double clock_end = MSG_get_clock();
+ double duration = clock_end - clock_sta;
+ double actual_speed = size / duration;
+#ifdef USE_MICRO_TASK
+ double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8;
+#else
+ double cpu_utilization = 0;
+#endif
+
+
+
if (stage == 2)
- XBT_INFO("mig-stage%d.%d: sent %f", stage, stage2_round, size);
+ XBT_INFO("mig-stage%d.%d: sent %f duration %f actual_speed %f (target %f) cpu %f", stage, stage2_round, size, duration, actual_speed, mig_speed, cpu_utilization);
else
- XBT_INFO("mig-stage%d: sent %f", stage, size);
+ XBT_INFO("mig-stage%d: sent %f duration %f actual_speed %f (target %f) cpu %f", stage, size, duration, actual_speed, mig_speed, cpu_utilization);
xbt_free(task_name);
+
+
+
+#ifdef USE_MICRO_TASK
+ /* The name of a micro task starts with __micro, which does not match the
+ * special name that finalizes the receiver loop. Thus, we send the special task.
+ **/
+ {
+ if (stage == 3) {
+ char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
+ msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
+ msg_error_t ret = MSG_task_send(task, mbox);
+ xbt_assert(ret == MSG_OK);
+ xbt_free(task_name);
+ }
+ }
+#endif
}
+double get_updated_size(double computed, double dp_rate, double dp_cap)
+{
+ double updated_size = computed * dp_rate;
+ XBT_INFO("updated_size %f dp_rate %f", updated_size, dp_rate);
+ if (updated_size > dp_cap) {
+ // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
+ updated_size = dp_cap;
+ }
-static int migration_tx_fun(int argc, char *argv[])
+ return updated_size;
+}
+
+static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *dst_pm_name,
+ long ramsize, double mig_speed, double xfer_cpu_overhead, double dp_rate, double dp_cap, double dpt_cpu_overhead)
{
- const char *pr_name = MSG_process_get_name(MSG_process_self());
- const char *host_name = MSG_host_get_name(MSG_host_self());
+ const char *vm_name = MSG_host_get_name(vm);
+ char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
+
+ const long chunksize = 1024 * 1024 * 100;
+ long remaining = ramsize;
+ double computed_total = 0;
+
+ while (remaining > 0) {
+ long datasize = chunksize;
+ if (remaining < chunksize)
+ datasize = remaining;
+
+ remaining -= datasize;
+
+ send_migration_data(vm_name, src_pm_name, dst_pm_name, datasize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
+
+ double computed = lookup_computed_flop_counts(vm, 1, 0);
+ computed_total += computed;
+
+ {
+ double updated_size = get_updated_size(computed, dp_rate, dp_cap);
+
+ double overhead = dpt_cpu_overhead * updated_size;
+ launch_deferred_exec_process(vm, overhead, 10000);
+ }
+ }
+
+ return computed_total;
+}
+
+
+
+static int migration_tx_fun(int argc, char *argv[])
+{
XBT_DEBUG("mig: tx_start");
xbt_assert(argc == 4);
simcall_host_get_params(vm, ¶ms);
const long ramsize = params.ramsize;
const long devsize = params.devsize;
+ const int skip_stage1 = params.skip_stage1;
const int skip_stage2 = params.skip_stage2;
- const double max_downtime = params.max_downtime;
const double dp_rate = params.dp_rate;
const double dp_cap = params.dp_cap;
const double mig_speed = params.mig_speed;
+ const double xfer_cpu_overhead = params.xfer_cpu_overhead;
+ const double dpt_cpu_overhead = params.dpt_cpu_overhead;
+
double remaining_size = ramsize + devsize;
- double threshold = max_downtime * 125 * 1000 * 1000;
+
+ double max_downtime = params.max_downtime;
+ if (max_downtime == 0) {
+ XBT_WARN("use the default max_downtime value 30ms");
+ max_downtime = 0.03;
+ }
+
+ /* This value assumes the network link is 1Gbps. */
+ double threshold = max_downtime * 125 * 1024 * 1024;
+
+ /* setting up parameters has done */
if (ramsize == 0)
/* Stage1: send all memory pages to the destination. */
start_dirty_page_tracking(vm);
- send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed);
-
- remaining_size -= ramsize;
+ double computed_during_stage1 = 0;
+ if (!skip_stage1) {
+ // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
+ /* send ramsize, but split it */
+ computed_during_stage1 = send_stage1(vm, src_pm_name, dst_pm_name, ramsize, mig_speed, xfer_cpu_overhead, dp_rate, dp_cap, dpt_cpu_overhead);
+ remaining_size -= ramsize;
+ }
/* Stage2: send update pages iteratively until the size of remaining states
int stage2_round = 0;
for (;;) {
- // long updated_size = lookup_dirty_pages(vm);
- double updated_size = lookup_computed_flop_counts(vm, stage2_round) * dp_rate;
- if (updated_size > dp_cap) {
- XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f",
- stage2_round, updated_size, dp_cap);
- updated_size = dp_cap;
+
+ double updated_size = 0;
+ if (stage2_round == 0) {
+ /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
+ updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
+ } else {
+ double computed = lookup_computed_flop_counts(vm, 2, stage2_round);
+ updated_size = get_updated_size(computed, dp_rate, dp_cap);
}
- remaining_size += updated_size;
+ XBT_INFO("%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f",
+ stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
- XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
- remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
- if (remaining_size < threshold)
- break;
+ if (stage2_round != 0) {
+ /* during stage1, we have already created overhead tasks */
+ double overhead = dpt_cpu_overhead * updated_size;
+ XBT_INFO("updated %f overhead %f", updated_size, overhead);
+ launch_deferred_exec_process(vm, overhead, 10000);
+ }
+
+
+ {
+ remaining_size += updated_size;
+
+ XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
+ remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
+
+ if (remaining_size < threshold)
+ break;
+ }
+
- send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed);
+ send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead);
remaining_size -= updated_size;
stage2_round += 1;
simcall_vm_suspend(vm);
stop_dirty_page_tracking(vm);
- send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed);
+ send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed, xfer_cpu_overhead);
xbt_free(mbox);
argv[3] = xbt_strdup(sg_host_name(dst_pm));
argv[4] = NULL;
- msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
+ MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
xbt_free(pr_name);
}
argv[2] = xbt_strdup(sg_host_name(src_pm));
argv[3] = xbt_strdup(sg_host_name(dst_pm));
argv[4] = NULL;
- msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
+ MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
xbt_free(pr_name);
}