Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
Merge branch 'task-dispatch' into 'master'
authorMartin Quinson <martin.quinson@ens-rennes.fr>
Wed, 11 Oct 2023 10:35:09 +0000 (10:35 +0000)
committerMartin Quinson <martin.quinson@ens-rennes.fr>
Wed, 11 Oct 2023 10:35:09 +0000 (10:35 +0000)
Task dispatch

See merge request simgrid/simgrid!175

17 files changed:
MANIFEST.in
docs/source/app_s4u.rst
examples/cpp/CMakeLists.txt
examples/cpp/task-dispatch/s4u-task-dispatch.cpp [new file with mode: 0644]
examples/cpp/task-dispatch/s4u-task-dispatch.tesh [new file with mode: 0644]
examples/cpp/task-parallelism/s4u-task-parallelism.tesh
examples/cpp/task-storm/s4u-task-storm.cpp
examples/cpp/task-storm/s4u-task-storm.tesh
examples/cpp/task-switch-host/s4u-task-switch-host.cpp
examples/python/task-io/task-io.py
examples/python/task-simple/task-simple.py
examples/python/task-simple/task-simple.tesh
examples/python/task-switch-host/task-switch-host.py
examples/python/task-variable-load/task-variable-load.py
include/simgrid/s4u/Task.hpp
src/bindings/python/simgrid_python.cpp
src/s4u/s4u_Task.cpp

index d33f19f..1b8ad27 100644 (file)
@@ -394,6 +394,8 @@ include examples/cpp/synchro-mutex/s4u-synchro-mutex.tesh
 include examples/cpp/synchro-semaphore/s4u-mc-synchro-semaphore.tesh
 include examples/cpp/synchro-semaphore/s4u-synchro-semaphore.cpp
 include examples/cpp/synchro-semaphore/s4u-synchro-semaphore.tesh
+include examples/cpp/task-dispatch/s4u-task-dispatch.cpp
+include examples/cpp/task-dispatch/s4u-task-dispatch.tesh
 include examples/cpp/task-io/s4u-task-io.cpp
 include examples/cpp/task-io/s4u-task-io.tesh
 include examples/cpp/task-microservice/s4u-task-microservice.cpp
index 23e0e04..8695720 100644 (file)
@@ -203,32 +203,33 @@ Repeatable Activities
 
 In order to simulate the execution of Dataflow applications, we introduced the
 concept of |API_s4u_Tasks|, that can be seen as repeatable activities. A Dataflow
-is defined as a graph of |API_s4u_Tasks| through which circulate Tokens. Tokens
-can carry any user-defined data, using the same internal mechanisms as for the
-other simulated objects. Each Task has to receive a token from each of its
-predecessor to fire a new instance of a :ref:`Communication <API_s4u_Comm>`,
-:ref:`Execution <API_s4u_Exec>`, or :ref:`I/O <API_s4u_Io>` activity.
-On completion of this activity, the Task propagates tokens
-to its successors, and waits for the next set of tokens to arrive.
-Multiple instances of the same Task can run in parallel by adjusting its
-horizontal scaling with
-:cpp:func:`s4u::Task::set_parallelism_degree() <simgrid::s4u::Task::set_parallelism_degree>`.
+is defined as a graph of |API_s4u_Tasks|, where each |API_s4u_Tasks| has a set of
+successors and predecessors. When a |API_s4u_Tasks| ends it sends a token to each
+of its successors. Each |API_s4u_Tasks| has to receive a token from each of its
+predecessor to start. Tokens can carry any user-defined data.
 
-:ref:`Communications <API_s4u_Comm>` (started on Mailboxes and consuming links),
-:ref:`Executions <API_s4u_Exec>` (started on Host and consuming CPU resources)
-:ref:`I/O <API_s4u_Io>` (started on and consuming disks).
+|API_s4u_Tasks| are composed of several instances: a dispatcher, a collector, and
+instance_0 to instance_n. The dispatcher rely on a load balancing function to select
+the next instance to fire. Once this instance finishes it fires the collector.
+
+Each instance of an |API_s4u_ExecTask| can be placed on a different host.
+|API_s4u_Comm| activities are automatically created when an instance triggers
+another instance on a different host. Each instance has its own parallelism degree
+to scale horizontally on several cores.
 
 To initiate the execution of a Dataflow, it is possible to some make
 |API_s4u_Tasks| fire one or more activities without waiting for any token with the
 :cpp:func:`s4u::Task::enqueue_firings() <simgrid::s4u::Task::enqueue_firings>`
 function.
 
-The parameters and successors of a Task can be redefined at runtime by attaching
+The parameters of Tasks can be redefined at runtime by attaching
 callbacks to the
 :cpp:func:`s4u::Task::on_this_start <simgrid::s4u::Task::on_this_start>`
 and
 :cpp:func:`s4u::Task::on_this_completion <simgrid::s4u::Task::on_this_completion>`
-signals.
+signals. The former is triggered by instances others than the dispatcher and the collector,
+and the latter is triggered by the collector.
+
 
 
 .. _s4u_mailbox:
index 2ddf038..b917540 100644 (file)
@@ -172,7 +172,7 @@ foreach (example activityset-testany activityset-waitany activityset-waitall act
                  mc-bugged1 mc-bugged1-liveness mc-bugged2 mc-bugged2-liveness mc-centralized-mutex mc-electric-fence mc-failing-assert
                  network-ns3 network-ns3-wifi network-wifi
                  io-async io-priority io-degradation io-file-system io-file-remote io-disk-raw io-dependent
-                 task-io task-microservice task-parallelism task-simple task-storm task-switch-host task-variable-load 
+                 task-dispatch task-io task-microservice task-parallelism task-simple task-storm task-switch-host task-variable-load
                  solar-panel-simple
                  platform-comm-serialize platform-failures platform-profile platform-properties
                  plugin-host-load plugin-jbod plugin-link-load plugin-prodcons
diff --git a/examples/cpp/task-dispatch/s4u-task-dispatch.cpp b/examples/cpp/task-dispatch/s4u-task-dispatch.cpp
new file mode 100644 (file)
index 0000000..10bd0d2
--- /dev/null
@@ -0,0 +1,103 @@
+/* Copyright (c) 2017-2023. 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 "simgrid/s4u.hpp"
+
+XBT_LOG_NEW_DEFAULT_CATEGORY(task_dispatch, "Messages specific for this s4u example");
+namespace sg4 = simgrid::s4u;
+
+static void manager(sg4::ExecTaskPtr t)
+{
+  auto PM0 = sg4::Engine::get_instance()->host_by_name("PM0");
+  auto PM1 = sg4::Engine::get_instance()->host_by_name("PM1");
+
+  XBT_INFO("Test set_flops");
+  t->enqueue_firings(2);
+  sg4::this_actor::sleep_for(50);
+  XBT_INFO("Set instance_0 flops to 50.");
+  t->set_flops(50 * PM0->get_speed());
+  sg4::this_actor::sleep_for(250);
+  t->set_flops(100 * PM0->get_speed());
+
+  XBT_INFO("Test set_parallelism degree");
+  t->enqueue_firings(3);
+  sg4::this_actor::sleep_for(50);
+  XBT_INFO("Set Task parallelism degree to 2.");
+  t->set_parallelism_degree(2);
+  sg4::this_actor::sleep_for(250);
+  t->set_parallelism_degree(1);
+
+  XBT_INFO("Test set_host dispatcher");
+  t->enqueue_firings(2);
+  sg4::this_actor::sleep_for(50);
+  XBT_INFO("Move dispatcher to PM1");
+  t->set_host(PM1, "dispatcher");
+  t->set_internal_bytes(1e6, "dispatcher");
+  sg4::this_actor::sleep_for(250);
+  t->set_host(PM0, "dispatcher");
+
+  XBT_INFO("Test set_host instance_0");
+  t->enqueue_firings(2);
+  sg4::this_actor::sleep_for(50);
+  XBT_INFO("Move instance_0 to PM1");
+  t->set_host(PM1, "instance_0");
+  t->set_flops(100 * PM1->get_speed());
+  t->set_internal_bytes(1e6, "instance_0");
+  sg4::this_actor::sleep_for(250);
+  t->set_host(PM0, "instance_0");
+  t->set_flops(100 * PM0->get_speed());
+
+  XBT_INFO("Test set_host collector");
+  t->enqueue_firings(2);
+  sg4::this_actor::sleep_for(50);
+  XBT_INFO("Move collector to PM1");
+  t->set_host(PM1, "collector");
+  sg4::this_actor::sleep_for(250);
+  t->set_host(PM0, "collector");
+
+  XBT_INFO("Test add_instances");
+  t->enqueue_firings(1);
+  sg4::this_actor::sleep_for(50);
+  XBT_INFO("Add 1 instance and update load balancing function");
+  t->add_instances(1);
+  t->set_load_balancing_function([]() {
+    static int round_robin_counter = 0;
+    int ret                        = round_robin_counter;
+    round_robin_counter            = round_robin_counter == 1 ? 0 : round_robin_counter + 1;
+    return "instance_" + std::to_string(ret);
+  });
+  t->enqueue_firings(2);
+  sg4::this_actor::sleep_for(250);
+
+  XBT_INFO("Test remove_instances");
+  XBT_INFO("Remove 1 instance and update load balancing function");
+  t->remove_instances(1);
+  t->set_load_balancing_function([]() { return "instance_0"; });
+  t->enqueue_firings(2);
+  sg4::this_actor::sleep_for(300);
+}
+
+int main(int argc, char* argv[])
+{
+  sg4::Engine e(&argc, argv);
+  e.load_platform(argv[1]);
+  auto PM0 = e.host_by_name("PM0");
+  auto PM1 = sg4::Engine::get_instance()->host_by_name("PM1");
+
+  auto a = sg4::ExecTask::init("A", 100 * PM0->get_speed(), PM0);
+  auto b = sg4::ExecTask::init("B", 50 * PM0->get_speed(), PM0);
+  auto c = sg4::CommTask::init("C", 1e6, PM1, PM0);
+
+  a->add_successor(b);
+
+  sg4::Task::on_completion_cb(
+      [](const sg4::Task* t) { XBT_INFO("Task %s finished (%d)", t->get_name().c_str(), t->get_count()); });
+  sg4::Task::on_start_cb([](const sg4::Task* t) { XBT_INFO("Task %s start", t->get_name().c_str()); });
+
+  sg4::Actor::create("manager", PM0, manager, a);
+
+  e.run();
+  return 0;
+}
diff --git a/examples/cpp/task-dispatch/s4u-task-dispatch.tesh b/examples/cpp/task-dispatch/s4u-task-dispatch.tesh
new file mode 100644 (file)
index 0000000..a4ace54
--- /dev/null
@@ -0,0 +1,82 @@
+#!/usr/bin/env tesh
+
+> > $ ${bindir:=.}/s4u-task-parallelism ${platfdir}/three_multicore_hosts.xml
+> [PM0:manager:(1) 0.000000] [task_dispatch/INFO] Test set_flops
+> [0.000000] [task_dispatch/INFO] Task A start
+> [0.000000] [task_dispatch/INFO] Task A start
+> [PM0:manager:(1) 50.000000] [task_dispatch/INFO] Set instance_0 flops to 50.
+> [100.000000] [task_dispatch/INFO] Task A finished (1)
+> [100.000000] [task_dispatch/INFO] Task B start
+> [150.000000] [task_dispatch/INFO] Task A finished (2)
+> [150.000000] [task_dispatch/INFO] Task B start
+> [150.000000] [task_dispatch/INFO] Task B finished (1)
+> [200.000000] [task_dispatch/INFO] Task B finished (2)
+> [PM0:manager:(1) 300.000000] [task_dispatch/INFO] Test set_parallelism degree
+> [300.000000] [task_dispatch/INFO] Task A start
+> [300.000000] [task_dispatch/INFO] Task A start
+> [300.000000] [task_dispatch/INFO] Task A start
+> [PM0:manager:(1) 350.000000] [task_dispatch/INFO] Set Task parallelism degree to 2.
+> [400.000000] [task_dispatch/INFO] Task A finished (3)
+> [400.000000] [task_dispatch/INFO] Task B start
+> [450.000000] [task_dispatch/INFO] Task A finished (4)
+> [450.000000] [task_dispatch/INFO] Task B start
+> [450.000000] [task_dispatch/INFO] Task B finished (3)
+> [500.000000] [task_dispatch/INFO] Task A finished (5)
+> [500.000000] [task_dispatch/INFO] Task B start
+> [500.000000] [task_dispatch/INFO] Task B finished (4)
+> [550.000000] [task_dispatch/INFO] Task B finished (5)
+> [PM0:manager:(1) 600.000000] [task_dispatch/INFO] Test set_host dispatcher
+> [600.000000] [task_dispatch/INFO] Task A start
+> [600.000000] [task_dispatch/INFO] Task A start
+> [PM0:manager:(1) 650.000000] [task_dispatch/INFO] Move dispatcher to PM1
+> [700.000000] [task_dispatch/INFO] Task A finished (6)
+> [700.000000] [task_dispatch/INFO] Task B start
+> [750.000000] [task_dispatch/INFO] Task B finished (6)
+> [800.009961] [task_dispatch/INFO] Task A finished (7)
+> [800.009961] [task_dispatch/INFO] Task B start
+> [850.009961] [task_dispatch/INFO] Task B finished (7)
+> [PM0:manager:(1) 900.000000] [task_dispatch/INFO] Test set_host instance_0
+> [900.000000] [task_dispatch/INFO] Task A start
+> [900.000000] [task_dispatch/INFO] Task A start
+> [PM0:manager:(1) 950.000000] [task_dispatch/INFO] Move instance_0 to PM1
+> [1000.000000] [task_dispatch/INFO] Task A finished (8)
+> [1000.000000] [task_dispatch/INFO] Task B start
+> [1050.000000] [task_dispatch/INFO] Task B finished (8)
+> [1100.019922] [task_dispatch/INFO] Task A finished (9)
+> [1100.019922] [task_dispatch/INFO] Task B start
+> [1150.019922] [task_dispatch/INFO] Task B finished (9)
+> [PM0:manager:(1) 1200.000000] [task_dispatch/INFO] Test set_host collector
+> [1200.000000] [task_dispatch/INFO] Task A start
+> [1200.000000] [task_dispatch/INFO] Task A start
+> [PM0:manager:(1) 1250.000000] [task_dispatch/INFO] Move collector to PM1
+> [1300.000000] [task_dispatch/INFO] Task A finished (10)
+> [1300.000000] [task_dispatch/INFO] Task B start
+> [1350.000000] [task_dispatch/INFO] Task B finished (10)
+> [1400.009961] [task_dispatch/INFO] Task A finished (11)
+> [1400.009961] [task_dispatch/INFO] Task B start
+> [1450.009961] [task_dispatch/INFO] Task B finished (11)
+> [PM0:manager:(1) 1500.000000] [task_dispatch/INFO] Test add_instances
+> [1500.000000] [task_dispatch/INFO] Task A start
+> [PM0:manager:(1) 1550.000000] [task_dispatch/INFO] Add 1 instance and update load balancing function
+> [1550.000000] [task_dispatch/INFO] Task A start
+> [1550.000000] [task_dispatch/INFO] Task A start
+> [1600.000000] [task_dispatch/INFO] Task A finished (12)
+> [1600.000000] [task_dispatch/INFO] Task B start
+> [1650.000000] [task_dispatch/INFO] Task A finished (13)
+> [1650.000000] [task_dispatch/INFO] Task B start
+> [1650.000000] [task_dispatch/INFO] Task B finished (12)
+> [1700.000000] [task_dispatch/INFO] Task A finished (14)
+> [1700.000000] [task_dispatch/INFO] Task B start
+> [1700.000000] [task_dispatch/INFO] Task B finished (13)
+> [1750.000000] [task_dispatch/INFO] Task B finished (14)
+> [PM0:manager:(1) 1800.000000] [task_dispatch/INFO] Test remove_instances
+> [PM0:manager:(1) 1800.000000] [task_dispatch/INFO] Remove 1 instance and update load balancing function
+> [1800.000000] [task_dispatch/INFO] Task A start
+> [1800.000000] [task_dispatch/INFO] Task A start
+> [1900.000000] [task_dispatch/INFO] Task A finished (15)
+> [1900.000000] [task_dispatch/INFO] Task B start
+> [1950.000000] [task_dispatch/INFO] Task B finished (15)
+> [2000.000000] [task_dispatch/INFO] Task A finished (16)
+> [2000.000000] [task_dispatch/INFO] Task B start
+> [2050.000000] [task_dispatch/INFO] Task B finished (16)
+> 
\ No newline at end of file
index 938ad96..492d015 100644 (file)
@@ -2,40 +2,40 @@
 
 $ ${bindir:=.}/s4u-task-parallelism ${platfdir}/three_multicore_hosts.xml
 > [0.000000] [task_parallelism/INFO] Task exec_A start
-> [100.000000] [task_parallelism/INFO] Task exec_A finished (1)
 > [100.000000] [task_parallelism/INFO] Task exec_A start
+> [100.000000] [task_parallelism/INFO] Task exec_A finished (1)
 > [200.000000] [task_parallelism/INFO] Task exec_A finished (2)
 > [300.000000] [task_parallelism/INFO] Task exec_A start
 > [300.000000] [task_parallelism/INFO] Task exec_A start
-> [400.000000] [task_parallelism/INFO] Task exec_A finished (3)
 > [400.000000] [task_parallelism/INFO] Task exec_A start
-> [400.000000] [task_parallelism/INFO] Task exec_A finished (4)
 > [400.000000] [task_parallelism/INFO] Task exec_A start
+> [400.000000] [task_parallelism/INFO] Task exec_A finished (3)
+> [400.000000] [task_parallelism/INFO] Task exec_A finished (4)
 > [500.000000] [task_parallelism/INFO] Task exec_A finished (5)
 > [500.000000] [task_parallelism/INFO] Task exec_A finished (6)
 > [600.000000] [task_parallelism/INFO] Task exec_A start
-> [700.000000] [task_parallelism/INFO] Task exec_A finished (7)
 > [700.000000] [task_parallelism/INFO] Task exec_A start
+> [700.000000] [task_parallelism/INFO] Task exec_A finished (7)
 > [800.000000] [task_parallelism/INFO] Task exec_A finished (8)
 > [900.000000] [task_parallelism/INFO] Task exec_A start
 > [900.000000] [task_parallelism/INFO] Task exec_A start
-> [1000.000000] [task_parallelism/INFO] Task exec_A finished (9)
 > [1000.000000] [task_parallelism/INFO] Task exec_A start
-> [1000.000000] [task_parallelism/INFO] Task exec_A finished (10)
 > [1000.000000] [task_parallelism/INFO] Task exec_A start
+> [1000.000000] [task_parallelism/INFO] Task exec_A finished (9)
+> [1000.000000] [task_parallelism/INFO] Task exec_A finished (10)
+> [1100.000000] [task_parallelism/INFO] Task exec_A start
 > [1100.000000] [task_parallelism/INFO] Task exec_A finished (11)
 > [1100.000000] [task_parallelism/INFO] Task exec_A finished (12)
-> [1100.000000] [task_parallelism/INFO] Task exec_A start
-> [1200.000000] [task_parallelism/INFO] Task exec_A finished (13)
 > [1200.000000] [task_parallelism/INFO] Task exec_A start
+> [1200.000000] [task_parallelism/INFO] Task exec_A finished (13)
 > [1250.000000] [task_parallelism/INFO] Task exec_A start
 > [1250.000000] [task_parallelism/INFO] Task exec_A start
-> [1300.000000] [task_parallelism/INFO] Task exec_A finished (14)
 > [1300.000000] [task_parallelism/INFO] Task exec_A start
-> [1350.000000] [task_parallelism/INFO] Task exec_A finished (15)
+> [1300.000000] [task_parallelism/INFO] Task exec_A finished (14)
 > [1350.000000] [task_parallelism/INFO] Task exec_A start
-> [1350.000000] [task_parallelism/INFO] Task exec_A finished (16)
 > [1350.000000] [task_parallelism/INFO] Task exec_A start
+> [1350.000000] [task_parallelism/INFO] Task exec_A finished (15)
+> [1350.000000] [task_parallelism/INFO] Task exec_A finished (16)
 > [1400.000000] [task_parallelism/INFO] Task exec_A finished (17)
 > [1450.000000] [task_parallelism/INFO] Task exec_A finished (18)
 > [1450.000000] [task_parallelism/INFO] Task exec_A finished (19)
\ No newline at end of file
index 58d08c3..0a0ab71 100644 (file)
@@ -74,10 +74,10 @@ int main(int argc, char* argv[])
      Alternatively we: remove/add the link between SA and SA_to_B2
                        add/remove the link between SA and SA_to_B1
   */
-  SA->on_this_start_cb([SA_to_B1, SA_to_B2](sg4::Task* t) {
+  SA->on_this_completion_cb([&SA_to_B1, &SA_to_B2](sg4::Task* t) {
     int count = t->get_count();
     sg4::CommTaskPtr comm;
-    if (count % 2 == 0) {
+    if (count % 2 == 1) {
       t->remove_successor(SA_to_B2);
       t->add_successor(SA_to_B1);
       comm = SA_to_B1;
@@ -86,7 +86,8 @@ int main(int argc, char* argv[])
       t->add_successor(SA_to_B2);
       comm = SA_to_B2;
     }
-    std::vector<double> amount = {1e3, 1e6, 1e9};
+    std::vector<double> amount = {1e9, 1e3, 1e6};
+    // XBT_INFO("Comm %f", amount[count % 3]);
     comm->set_amount(amount[count % 3]);
     auto token = std::make_shared<sg4::Token>();
     token->set_data(new double(amount[count % 3]));
@@ -94,18 +95,26 @@ int main(int argc, char* argv[])
   });
 
   // The token sent by SA is forwarded by both communication tasks
-  SA_to_B1->on_this_start_cb([&SA](sg4::Task* t) { t->set_token(t->get_next_token_from(SA)); });
-  SA_to_B2->on_this_start_cb([&SA](sg4::Task* t) { t->set_token(t->get_next_token_from(SA)); });
+  SA_to_B1->on_this_completion_cb([&SA](sg4::Task* t) {
+    t->set_token(t->get_token_from(SA));
+    t->deque_token_from(SA);
+  });
+  SA_to_B2->on_this_completion_cb([&SA](sg4::Task* t) {
+    t->set_token(t->get_token_from(SA));
+    t->deque_token_from(SA);
+  });
 
   /* B1 and B2 read the value of the token received by their predecessors
      and use it to adapt their amount of work to do.
   */
-  B1->on_this_start_cb([SA_to_B1](sg4::Task* t) {
-    auto data = t->get_next_token_from(SA_to_B1)->get_unique_data<double>();
+  B1->on_this_start_cb([&SA_to_B1](sg4::Task* t) {
+    auto data = t->get_token_from(SA_to_B1)->get_data<double>();
+    t->deque_token_from(SA_to_B1);
     t->set_amount(*data * 10);
   });
-  B2->on_this_start_cb([SA_to_B2](sg4::Task* t) {
-    auto data = t->get_next_token_from(SA_to_B2)->get_unique_data<double>();
+  B2->on_this_start_cb([&SA_to_B2](sg4::Task* t) {
+    auto data = t->get_token_from(SA_to_B2)->get_data<double>();
+    t->deque_token_from(SA_to_B2);
     t->set_amount(*data * 10);
   });
 
index d7c364a..376dc31 100644 (file)
@@ -24,11 +24,11 @@ $ ${bindir:=.}/s4u-task-storm ${platfdir}/small_platform.xml
 > [1.798442] [task_storm/INFO] Task SB_to_B3 finished (5)
 > [2.619232] [task_storm/INFO] Task B3 finished (1)
 > [6.743624] [task_storm/INFO] Task B3 finished (2)
-> [10.868015] [task_storm/INFO] Task B3 finished (3)
 > [10.868015] [task_storm/INFO] Task B4 finished (1)
+> [10.868015] [task_storm/INFO] Task B3 finished (3)
 > [14.992407] [task_storm/INFO] Task B3 finished (4)
-> [19.116799] [task_storm/INFO] Task B3 finished (5)
 > [19.116799] [task_storm/INFO] Task B4 finished (2)
+> [19.116799] [task_storm/INFO] Task B3 finished (5)
 > [23.241190] [task_storm/INFO] Task B4 finished (3)
 > [27.365582] [task_storm/INFO] Task B4 finished (4)
 > [31.489974] [task_storm/INFO] Task B4 finished (5)
index b007523..c236bdb 100644 (file)
@@ -54,12 +54,19 @@ int main(int argc, char* argv[])
   // successors to comm0
   comm0->on_this_start_cb([&comm0, exec1, exec2, jupiter, fafard](const sg4::Task*) {
     static int count = 0;
-    if (count % 2 == 0) {
+    if (count % 2 == 0)
       comm0->set_destination(jupiter);
+    else
+      comm0->set_destination(fafard);
+    count++;
+  });
+
+  comm0->on_this_completion_cb([&comm0, exec1, exec2](const sg4::Task*) {
+    static int count = 0;
+    if (count % 2 == 0) {
       comm0->add_successor(exec1);
       comm0->remove_successor(exec2);
     } else {
-      comm0->set_destination(fafard);
       comm0->add_successor(exec2);
       comm0->remove_successor(exec1);
     }
index e75215a..d3ab8c1 100644 (file)
@@ -18,7 +18,7 @@ def parse():
     return parser.parse_args()
 
 def callback(t):
-    print(f'[{Engine.clock}] {t} finished ({t.count})')
+    print(f'[{Engine.clock}] {t} finished ({t.get_count()})')
 
 if __name__ == '__main__':
     args = parse()
index 23e9fc0..beca2b6 100644 (file)
@@ -28,7 +28,7 @@ def parse():
     return parser.parse_args()
 
 def callback(t):
-    print(f'[{Engine.clock}] {t} finished ({t.count})')
+    print(f'[{Engine.clock}] {t} finished ({t.get_count()})')
 
 if __name__ == '__main__':
     args = parse()
index 5a27a53..f9a828f 100644 (file)
@@ -5,5 +5,5 @@ $ ${pythoncmd:=python3} ${PYTHON_TOOL_OPTIONS:=} ${srcdir:=.}/task-simple.py --p
 > [11.714617112501687] CommTask(comm) finished (1)
 > [20.388399000968448] ExecTask(exec1) finished (2)
 > [21.90881661298591] CommTask(comm) finished (2)
-> [24.82146412938331] ExecTask(exec2) finished (1)
-> [37.92831114626493] ExecTask(exec2) finished (2)
+> [24.821464129383305] ExecTask(exec2) finished (1)
+> [37.928311146264925] ExecTask(exec2) finished (2)
index 5be8922..03dce6a 100644 (file)
@@ -44,12 +44,16 @@ def parse():
     return parser.parse_args()
 
 def callback(t):
-    print(f'[{Engine.clock}] {t} finished ({t.count})')
+    print(f'[{Engine.clock}] {t} finished ({t.get_count()})')
 
-def switch(t, hosts, execs):
-    comm0.destination = hosts[t.count % 2]
-    comm0.remove_successor(execs[t.count % 2 - 1])
-    comm0.add_successor(execs[t.count % 2])
+def switch_destination(t, hosts):
+    t.destination = hosts[switch_destination.count % 2]
+    switch_destination.count += 1
+switch_destination.count = 0
+
+def switch_successor(t, execs):
+    t.remove_successor(execs[t.get_count() % 2])
+    t.add_successor(execs[t.get_count() % 2 - 1])
 
 if __name__ == '__main__':
     args = parse()
@@ -74,13 +78,16 @@ if __name__ == '__main__':
     exec1.add_successor(comm1)
     exec2.add_successor(comm2)
 
-    # Add a function to be called when tasks end for log purpose
+    # Add a callback when tasks end for log purpose
     Task.on_completion_cb(callback)
 
-    # Add a function to be called before each firing of comm0
-    # This function modifies the graph of tasks by adding or removing
-    # successors to comm0
-    comm0.on_this_start_cb(lambda t: switch(t, [jupiter, fafard], [exec1,exec2]))
+    # Add a callback before each firing of comm0
+    # It switches the destination of comm0
+    comm0.on_this_start_cb(lambda t: switch_destination(t, [jupiter, fafard]))
+
+    # Add a callback before comm0 send tokens to successors
+    # It switches the successor of comm0
+    comm0.on_this_completion_cb(lambda t: switch_successor(t, [exec1,exec2]))
 
     # Enqueue two firings for task exec1
     comm0.enqueue_firings(4)
index 51dbc1a..14fc2ec 100644 (file)
@@ -28,7 +28,7 @@ def parse():
     return parser.parse_args()
 
 def callback(t):
-    print(f'[{Engine.clock}] {t} finished ({t.count})')
+    print(f'[{Engine.clock}] {t} finished ({t.get_count()})')
 
 def variable_load(t):
     print('--- Small load ---')
index f2b6eb6..03ee265 100644 (file)
@@ -10,6 +10,7 @@
 #include <map>
 #include <memory>
 #include <set>
+#include <xbt/asserts.h>
 
 namespace simgrid::s4u {
 
@@ -27,23 +28,29 @@ using IoTaskPtr = boost::intrusive_ptr<IoTask>;
 class XBT_PUBLIC Token : public xbt::Extendable<Token> {};
 
 class Task {
+
   std::string name_;
-  double amount_;
-  int queued_firings_     = 0;
-  int count_              = 0;
-  int running_instances_  = 0;
-  int parallelism_degree_ = 1;
+
+  std::map<std::string, double> amount_              = {{"instance_0", 0}, {"dispatcher", 0}, {"collector", 0}};
+  std::map<std::string, int> queued_firings_         = {{"instance_0", 0}, {"dispatcher", 0}, {"collector", 0}};
+  std::map<std::string, int> running_instances_      = {{"instance_0", 0}, {"dispatcher", 0}, {"collector", 0}};
+  std::map<std::string, int> count_                  = {{"instance_0", 0}, {"dispatcher", 0}, {"collector", 0}};
+  std::map<std::string, int> parallelism_degree_     = {{"instance_0", 1}, {"dispatcher", 1}, {"collector", 1}};
+  std::map<std::string, int> internal_bytes_to_send_ = {{"instance_0", 0}, {"dispatcher", 0}};
+
+  std::function<std::string()> load_balancing_function_;
 
   std::set<Task*> successors_                 = {};
   std::map<Task*, unsigned int> predecessors_ = {};
   std::atomic_int_fast32_t refcount_{0};
 
-  bool ready_to_run() const;
+  bool ready_to_run(std::string instance);
   void receive(Task* source);
 
   std::shared_ptr<Token> token_ = nullptr;
-  std::deque<std::map<TaskPtr, std::shared_ptr<Token>>> tokens_received_;
-  std::deque<ActivityPtr> current_activities_;
+  std::map<TaskPtr, std::deque<std::shared_ptr<Token>>> tokens_received_;
+  std::map<std::string, std::deque<ActivityPtr>> current_activities_ = {
+      {"instance_0", {}}, {"dispatcher", {}}, {"collector", {}}};
 
   inline static xbt::signal<void(Task*)> on_start;
   xbt::signal<void(Task*)> on_this_start;
@@ -54,23 +61,33 @@ protected:
   explicit Task(const std::string& name);
   virtual ~Task() = default;
 
-  virtual void fire();
-  void complete();
+  virtual void fire(std::string instance);
+  void complete(std::string instance);
 
-  void store_activity(ActivityPtr a) { current_activities_.push_back(a); }
+  void store_activity(ActivityPtr a, std::string instance) { current_activities_[instance].push_back(a); }
+
+  virtual void add_instances(int n);
+  virtual void remove_instances(int n);
 
 public:
   void set_name(std::string name);
   const std::string& get_name() const { return name_; }
   const char* get_cname() const { return name_.c_str(); }
-  void set_amount(double amount);
-  double get_amount() const { return amount_; }
-  int get_count() const { return count_; }
-  void set_parallelism_degree(int n);
-  int get_parallelism_degree() const { return parallelism_degree_; }
+  void set_amount(double amount, std::string instance = "instance_0");
+  double get_amount(std::string instance = "instance_0") const { return amount_.at(instance); }
+  int get_queued_firings(std::string instance = "instance_0") { return queued_firings_.at(instance); }
+  int get_running_count(std::string instance = "instance_0") { return running_instances_.at(instance); }
+  int get_count(std::string instance = "collector") const { return count_.at(instance); }
+  void set_parallelism_degree(int n, std::string instance = "all");
+  int get_parallelism_degree(std::string instance = "instance_0") const { return parallelism_degree_.at(instance); }
+  void set_internal_bytes(int bytes, std::string instance = "instance_0");
+  double get_internal_bytes(std::string instance = "instance_0") const { return internal_bytes_to_send_.at(instance); }
+  void set_load_balancing_function(std::function<std::string()> func);
 
   void set_token(std::shared_ptr<Token> token);
-  std::shared_ptr<Token> get_next_token_from(TaskPtr t);
+  std::shared_ptr<Token> get_token_from(TaskPtr t) const { return tokens_received_.at(t).front(); }
+  std::deque<std::shared_ptr<Token>> get_tokens_from(TaskPtr t) const { return tokens_received_.at(t); }
+  void deque_token_from(TaskPtr t);
 
   void add_successor(TaskPtr t);
   void remove_successor(TaskPtr t);
@@ -107,7 +124,7 @@ class CommTask : public Task {
   Host* destination_;
 
   explicit CommTask(const std::string& name);
-  void fire() override;
+  void fire(std::string instance) override;
 
 public:
   static CommTaskPtr init(const std::string& name);
@@ -118,30 +135,33 @@ public:
   CommTaskPtr set_destination(Host* destination);
   Host* get_destination() const { return destination_; }
   CommTaskPtr set_bytes(double bytes);
-  double get_bytes() const { return get_amount(); }
+  double get_bytes() const { return get_amount("instance_0"); }
 };
 
 class ExecTask : public Task {
-  Host* host_;
+  std::map<std::string, Host*> host_ = {{"instance_0", nullptr}, {"dispatcher", nullptr}, {"collector", nullptr}};
 
   explicit ExecTask(const std::string& name);
-  void fire() override;
+  void fire(std::string instance) override;
 
 public:
   static ExecTaskPtr init(const std::string& name);
   static ExecTaskPtr init(const std::string& name, double flops, Host* host);
 
-  ExecTaskPtr set_host(Host* host);
-  Host* get_host() const { return host_; }
-  ExecTaskPtr set_flops(double flops);
-  double get_flops() const { return get_amount(); }
+  ExecTaskPtr set_host(Host* host, std::string instance = "all");
+  Host* get_host(std::string instance = "instance_0") const { return host_.at(instance); }
+  ExecTaskPtr set_flops(double flops, std::string instance = "instance_0");
+  double get_flops(std::string instance = "instance_0") const { return get_amount(instance); }
+
+  void add_instances(int n) override;
+  void remove_instances(int n) override;
 };
 
 class IoTask : public Task {
   Disk* disk_;
   Io::OpType type_;
   explicit IoTask(const std::string& name);
-  void fire() override;
+  void fire(std::string instance) override;
 
 public:
   static IoTaskPtr init(const std::string& name);
@@ -150,7 +170,7 @@ public:
   IoTaskPtr set_disk(Disk* disk);
   Disk* get_disk() const { return disk_; }
   IoTaskPtr set_bytes(double bytes);
-  double get_bytes() const { return get_amount(); }
+  double get_bytes() const { return get_amount("instance_0"); }
   IoTaskPtr set_op_type(Io::OpType type);
   Io::OpType get_op_type() const { return type_; }
 };
index 792ae4b..2f2ef70 100644 (file)
@@ -861,9 +861,14 @@ PYBIND11_MODULE(simgrid, m)
           },
           "Add a callback called when each task ends.")
       .def_property_readonly("name", &Task::get_name, "The name of this task (read-only).")
-      .def_property_readonly("count", &Task::get_count, "The execution count of this task (read-only).")
       .def_property_readonly("successors", &Task::get_successors, "The successors of this task (read-only).")
       .def_property("amount", &Task::get_amount, &Task::set_amount, "The amount of work to do for this task.")
+      .def(
+          "get_count", [](const TaskPtr t) { return t->get_count("instance_0"); },
+          "The execution count of this task instance_0.")
+      .def(
+          "get_count", [](const TaskPtr t, const std::string& instance) { return t->get_count(instance); },
+          "The execution count of this task instance.")
       .def("enqueue_firings", py::overload_cast<int>(&Task::enqueue_firings), py::call_guard<py::gil_scoped_release>(),
            py::arg("n"), "Enqueue firings for this task.")
       .def("add_successor", py::overload_cast<TaskPtr>(&Task::add_successor), py::call_guard<py::gil_scoped_release>(),
index 7b04687..bcc2f4e 100644 (file)
@@ -1,45 +1,34 @@
+#include <cstddef>
 #include <memory>
 #include <simgrid/Exception.hpp>
+#include <simgrid/s4u/Activity.hpp>
 #include <simgrid/s4u/Comm.hpp>
+#include <simgrid/s4u/Disk.hpp>
 #include <simgrid/s4u/Exec.hpp>
 #include <simgrid/s4u/Io.hpp>
 #include <simgrid/s4u/Task.hpp>
 #include <simgrid/simix.hpp>
+#include <string>
+#include <xbt/asserts.h>
 
 #include "src/simgrid/module.hpp"
 
 SIMGRID_REGISTER_PLUGIN(task, "Battery management", nullptr)
-/**
-  @beginrst
-
-
-Tasks are designed to represent dataflows, i.e, graphs of Tasks.
-Tasks can only be instancied using either
-:cpp:func:`simgrid::s4u::ExecTask::init` or :cpp:func:`simgrid::s4u::CommTask::init`
-An ExecTask is an Execution Task. Its underlying Activity is an :ref:`Exec <API_s4u_Exec>`.
-A CommTask is a Communication Task. Its underlying Activity is a :ref:`Comm <API_s4u_Comm>`.
-
-
-
-  @endrst
- */
 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(Task, kernel, "Logging specific to the task plugin");
 
 namespace simgrid::s4u {
 
 Task::Task(const std::string& name) : name_(name) {}
 
-/**
- *  @brief Return True if the Task can start a new Activity.
- *  @note The Task is ready if not already doing something and there is at least one execution waiting in queue.
+/** @param instance The Task instance to check.
+ *  @brief Return True if this Task instance can start.
  */
-bool Task::ready_to_run() const
+bool Task::ready_to_run(std::string instance)
 {
-  return running_instances_ < parallelism_degree_ && queued_firings_ > 0;
+  return running_instances_[instance] < parallelism_degree_[instance] && queued_firings_[instance] > 0;
 }
 
-/**
- *  @param source The sender.
+/** @param source The sender.
  *  @brief Receive a token from another Task.
  *  @note Check upon reception if the Task has received a token from each of its predecessors,
  * and in this case consumes those tokens and enqueue an execution.
@@ -47,12 +36,10 @@ bool Task::ready_to_run() const
 void Task::receive(Task* source)
 {
   XBT_DEBUG("Task %s received a token from %s", name_.c_str(), source->name_.c_str());
-  auto source_count = predecessors_[source];
   predecessors_[source]++;
-  if (tokens_received_.size() <= queued_firings_ + source_count)
-    tokens_received_.emplace_back();
-  tokens_received_[queued_firings_ + source_count][source] = source->token_;
-  bool enough_tokens                                       = true;
+  if (source->token_ != nullptr)
+    tokens_received_[source].push_back(source->token_);
+  bool enough_tokens = true;
   for (auto const& [key, val] : predecessors_)
     if (val < 1) {
       enough_tokens = false;
@@ -65,54 +52,88 @@ void Task::receive(Task* source)
   }
 }
 
-/**
- *  @brief Task routine when finishing an execution.
- *  @note Set its working status as false.
- * Add 1 to its count of finished executions.
- * Call the on_this_end func.
- * Fire on_end callback.
- * Send a token to each of its successors.
- * Start a new execution if possible.
+/** @param instance The Taks instance to complete.
+ *  @brief Task instance routine when finishing an execution of an instance.
+ *  @note The dispatcher instance enqueues a firing for the next instance.
+ *        The collector instance triggers the on_completion signals and sends tokens to successors.
+ *        Others instances enqueue a firing of the collector instance.
  */
-void Task::complete()
+void Task::complete(std::string instance)
 {
   xbt_assert(Actor::is_maestro());
-  running_instances_--;
-  count_++;
-  on_this_completion(this);
-  on_completion(this);
-  for (auto const& t : successors_)
-    t->receive(this);
-  if (ready_to_run())
-    fire();
+  running_instances_[instance]--;
+  count_[instance]++;
+  if (instance == "collector") {
+    on_this_completion(this);
+    on_completion(this);
+    for (auto const& t : successors_)
+      t->receive(this);
+  } else if (instance == "dispatcher") {
+    auto next_instance = load_balancing_function_();
+    xbt_assert(next_instance != "dispatcher" and next_instance != "collector", "Invalid instance selected: %s",
+               next_instance.c_str());
+    queued_firings_[next_instance] = queued_firings_.at(next_instance) + 1;
+    while (ready_to_run(next_instance))
+      fire(next_instance);
+  } else {
+    queued_firings_["collector"]++;
+    while (ready_to_run("collector"))
+      fire("collector");
+  }
+  if (ready_to_run(instance))
+    fire(instance);
 }
 
-/** @param n The new parallelism degree of the Task.
- *  @brief Set the parallelism degree of the Task to inscrease or decrease horizontal scaling.
- *  @note When increasing the degree the function starts new instances if there is queued firings.
- *        When decreasing the degree the function does NOT stop running instances.
-
+/** @param n The new parallelism degree of the Task instance.
+ *  @param instance The Task instance to modify.
+ *  @note You can use instance "all" to modify the parallelism degree of all instances of this Task.
+ *        When increasing the degree new executions are started if there is queued firings.
+ *        When decreasing the degree instances already running are NOT stopped.
  */
-void Task::set_parallelism_degree(int n)
+void Task::set_parallelism_degree(int n, std::string instance)
 {
-  xbt_assert(n > 0, "Parallelism degree of Tasks must be above 0.");
-  simgrid::kernel::actor::simcall_answered([this, n] {
-    parallelism_degree_ = n;
-    while (ready_to_run())
-      fire();
+  xbt_assert(n > 0, "Parallelism degree must be above 0.");
+  simgrid::kernel::actor::simcall_answered([this, n, &instance] {
+    if (instance == "all") {
+      for (auto& [key, value] : parallelism_degree_) {
+        parallelism_degree_[key] = n;
+        while (ready_to_run(key))
+          fire(key);
+      }
+    } else {
+      parallelism_degree_[instance] = n;
+      while (ready_to_run(instance))
+        fire(instance);
+    }
   });
 }
 
+/** @param bytes The internal bytes of the Task instance.
+ *  @param instance The Task instance to modify.
+ *  @note Internal bytes are used for Comms between the dispatcher and instance_n,
+ *        and between instance_n and the collector if they are not on the same host.
+ */
+void Task::set_internal_bytes(int bytes, std::string instance)
+{
+  simgrid::kernel::actor::simcall_answered([this, bytes, &instance] { internal_bytes_to_send_[instance] = bytes; });
+}
+
+/** @param func The load balancing function.
+ *  @note The dispatcher uses this function to determine which instance to trigger next.
+ */
+void Task::set_load_balancing_function(std::function<std::string()> func)
+{
+  simgrid::kernel::actor::simcall_answered([this, func] { load_balancing_function_ = func; });
+}
+
 /** @param n The number of firings to enqueue.
- *  @brief Enqueue firing.
- *  @note Immediatly fire an activity if possible.
  */
 void Task::enqueue_firings(int n)
 {
   simgrid::kernel::actor::simcall_answered([this, n] {
-    queued_firings_ += n;
-    while (ready_to_run())
-      fire();
+    queued_firings_["dispatcher"] += n;
+    while (ready_to_run("dispatcher"))
+      fire("dispatcher");
   });
 }
 
@@ -125,46 +146,45 @@ void Task::set_name(std::string name)
 }
 
 /** @param amount The amount to set.
- *  @brief Set the amout of work to do.
+ *  @param instance The Task instance to modify.
  *  @note Amount in flop for ExecTask and in bytes for CommTask.
  */
-void Task::set_amount(double amount)
+void Task::set_amount(double amount, std::string instance)
 {
-  simgrid::kernel::actor::simcall_answered([this, amount] { amount_ = amount; });
+  simgrid::kernel::actor::simcall_answered([this, amount, &instance] { amount_[instance] = amount; });
 }
 
 /** @param token The token to set.
  *  @brief Set the token to send to successors.
- *  @note The token is passed to each successor after the task end, i.e., after the on_end callback.
+ *  @note The token is passed to each successor after the Task instance collector end, i.e., after the on_completion
+ * callback.
  */
 void Task::set_token(std::shared_ptr<Token> token)
 {
   simgrid::kernel::actor::simcall_answered([this, token] { token_ = token; });
 }
 
-/** @return Map of tokens received for the next execution.
- *  @note If there is no queued execution for this task the map might not exist or be partially empty.
+/** @param t The Task to deque a token from.
  */
-std::shared_ptr<Token> Task::get_next_token_from(TaskPtr t)
+void Task::deque_token_from(TaskPtr t)
 {
-  return tokens_received_.front()[t];
+  simgrid::kernel::actor::simcall_answered([this, &t] { tokens_received_[t].pop_front(); });
 }
 
-void Task::fire()
+void Task::fire(std::string instance)
 {
-  if ((int)current_activities_.size() > parallelism_degree_) {
-    current_activities_.pop_front();
+  if ((int)current_activities_[instance].size() > parallelism_degree_[instance]) {
+    current_activities_[instance].pop_front();
+  }
+  if (instance != "dispatcher" and instance != "collector") {
+    on_this_start(this);
+    on_start(this);
   }
-  on_this_start(this);
-  on_start(this);
-  running_instances_++;
-  queued_firings_ = std::max(queued_firings_ - 1, 0);
-  if (not tokens_received_.empty())
-    tokens_received_.pop_front();
+  running_instances_[instance]++;
+  queued_firings_[instance] = std::max(queued_firings_[instance] - 1, 0);
 }
 
-/** @param successor The Task to add.
- *  @brief Add a successor to this Task.
+/** @param successor The Task to add as a successor.
  *  @note It also adds this as a predecessor of successor.
  */
 void Task::add_successor(TaskPtr successor)
@@ -175,8 +195,7 @@ void Task::add_successor(TaskPtr successor)
   });
 }
 
-/** @param successor The Task to remove.
- *  @brief Remove a successor from this Task.
+/** @param successor The Task to remove from the successors of this Task.
  *  @note It also remove this from the predecessors of successor.
  */
 void Task::remove_successor(TaskPtr successor)
@@ -187,6 +206,8 @@ void Task::remove_successor(TaskPtr successor)
   });
 }
 
+/** @brief Remove all successors from this Task.
+ */
 void Task::remove_all_successors()
 {
   simgrid::kernel::actor::simcall_answered([this] {
@@ -198,12 +219,58 @@ void Task::remove_all_successors()
   });
 }
 
+/** @param n The number of instances to add to this Task (>=0).
+ *  @note Instances goes always from instance_0 to instance_x,
+ *        where x is the current number of instance.
+ */
+void Task::add_instances(int n)
+{
+  xbt_assert(n >= 0, "Cannot add a negative number of instances (provided: %d)", n);
+  int instance_count = (int)amount_.size() - 2;
+  for (int i = instance_count; i < n + instance_count; i++) {
+    amount_["instance_" + std::to_string(i)]                 = amount_.at("instance_0");
+    queued_firings_["instance_" + std::to_string(i)]         = 0;
+    running_instances_["instance_" + std::to_string(i)]      = 0;
+    count_["instance_" + std::to_string(i)]                  = 0;
+    parallelism_degree_["instance_" + std::to_string(i)]     = parallelism_degree_.at("instance_0");
+    current_activities_["instance_" + std::to_string(i)]     = {};
+    internal_bytes_to_send_["instance_" + std::to_string(i)] = internal_bytes_to_send_.at("instance_0");
+    ;
+  }
+}
+
+/** @param n The number of instances to remove from this Task (>=0).
+ *  @note Instances goes always from instance_0 to instance_x,
+ *        where x is the current number of instance.
+ *        Running instances cannot be removed.
+ */
+void Task::remove_instances(int n)
+{
+  int instance_count = (int)amount_.size() - 2;
+  xbt_assert(n >= 0, "Cannot remove a negative number of instances (provided: %d)", n);
+  xbt_assert(instance_count - n > 0, "The number of instances must be above 0 (instances: %d, provided: %d)",
+             instance_count, n);
+  for (int i = instance_count - 1; i >= instance_count - n; i--) {
+    xbt_assert(running_instances_.at("instance_" + std::to_string(i)) == 0,
+               "Cannot remove a running instance (instances: %d)", i);
+    amount_.erase("instance_" + std::to_string(i));
+    queued_firings_.erase("instance_" + std::to_string(i));
+    running_instances_.erase("instance_" + std::to_string(i));
+    count_.erase("instance_" + std::to_string(i));
+    parallelism_degree_.erase("instance_" + std::to_string(i));
+    current_activities_.erase("instance_" + std::to_string(i));
+  }
+}
+
 /**
  *  @brief Default constructor.
  */
-ExecTask::ExecTask(const std::string& name) : Task(name) {}
+ExecTask::ExecTask(const std::string& name) : Task(name)
+{
+  set_load_balancing_function([]() { return "instance_0"; });
+}
 
-/** @ingroup plugin_task
+/**
  *  @brief Smart Constructor.
  */
 ExecTaskPtr ExecTask::init(const std::string& name)
@@ -211,7 +278,7 @@ ExecTaskPtr ExecTask::init(const std::string& name)
   return ExecTaskPtr(new ExecTask(name));
 }
 
-/** @ingroup plugin_task
+/**
  *  @brief Smart Constructor.
  */
 ExecTaskPtr ExecTask::init(const std::string& name, double flops, Host* host)
@@ -219,45 +286,109 @@ ExecTaskPtr ExecTask::init(const std::string& name, double flops, Host* host)
   return init(name)->set_flops(flops)->set_host(host);
 }
 
-/**
- *  @brief Do one execution of the Task.
- *  @note Call the on_this_start() func.
- *  Init and start the underlying Activity.
+/** @param instance The Task instance to fire.
+ *  @note Only the dispatcher instance triggers the on_start signal.
+ *        Comms are created if hosts differ between dispatcher and the instance to fire,
+ *        or between the instance and the collector.
  */
-void ExecTask::fire()
+void ExecTask::fire(std::string instance)
 {
-  Task::fire();
-  auto exec = Exec::init()->set_name(get_name())->set_flops_amount(get_amount())->set_host(host_);
-  exec->start();
-  exec->on_this_completion_cb([this](Exec const&) { complete(); });
-  store_activity(exec);
+  Task::fire(instance);
+  if (instance == "dispatcher" or instance == "collector") {
+    auto exec = Exec::init()
+                    ->set_name(get_name() + "_" + instance)
+                    ->set_flops_amount(get_amount(instance))
+                    ->set_host(host_[instance]);
+    exec->start();
+    exec->on_this_completion_cb([this, instance](Exec const&) { complete(instance); });
+    store_activity(exec, instance);
+  } else {
+    auto exec = Exec::init()->set_name(get_name())->set_flops_amount(get_amount())->set_host(host_[instance]);
+    if (host_["dispatcher"] == host_[instance]) {
+      exec->start();
+      store_activity(exec, instance);
+    } else {
+      auto comm = Comm::sendto_init(host_["dispatcher"], host_[instance])
+                      ->set_name(get_name() + "_dispatcher_to_" + instance)
+                      ->set_payload_size(get_internal_bytes("dispatcher"));
+      comm->add_successor(exec);
+      comm->start();
+      store_activity(comm, instance);
+    }
+    if (host_[instance] == host_["collector"]) {
+      exec->on_this_completion_cb([this, instance](Exec const&) { complete(instance); });
+      if (host_["dispatcher"] != host_[instance])
+        store_activity(exec, instance);
+    } else {
+      auto comm = Comm::sendto_init(host_[instance], host_["collector"])
+                      ->set_name(get_name() + instance + "_to_collector")
+                      ->set_payload_size(get_internal_bytes(instance));
+      exec->add_successor(comm);
+      comm->on_this_completion_cb([this, instance](Comm const&) { complete(instance); });
+      comm.detach();
+    }
+  }
 }
 
-/** @ingroup plugin_task
- *  @param host The host to set.
+/** @param host The host to set.
+ *  @param instance The Task instance to modify.
  *  @brief Set a new host.
  */
-ExecTaskPtr ExecTask::set_host(Host* host)
+ExecTaskPtr ExecTask::set_host(Host* host, std::string instance)
 {
-  kernel::actor::simcall_answered([this, host] { host_ = host; });
+  kernel::actor::simcall_answered([this, host, &instance] {
+    if (instance == "all")
+      for (auto& [key, value] : host_)
+        host_[key] = host;
+    else
+      host_[instance] = host;
+  });
   return this;
 }
 
-/** @ingroup plugin_task
- *  @param flops The amount of flops to set.
+/** @param flops The amount of flops to set.
+ *  @param instance The Task instance to modify.
  */
-ExecTaskPtr ExecTask::set_flops(double flops)
+ExecTaskPtr ExecTask::set_flops(double flops, std::string instance)
 {
-  kernel::actor::simcall_answered([this, flops] { set_amount(flops); });
+  kernel::actor::simcall_answered([this, flops, &instance] { set_amount(flops, instance); });
   return this;
 }
 
+/** @param n The number of instances to add to this Task (>=0).
+    @note Instances goes always from instance_0 to instance_x,
+          where x is the current number of instance.
+ */
+void ExecTask::add_instances(int n)
+{
+  Task::add_instances(n);
+  int instance_count = (int)host_.size() - 2;
+  for (int i = instance_count; i < n + instance_count; i++)
+    host_["instance_" + std::to_string(i)] = host_.at("instance_0");
+}
+
+/** @param n The number of instances to remove from this Task (>=0).
+    @note Instances goes always from instance_0 to instance_x,
+          where x is the current number of instance.
+          Running instance cannot be removed.
+ */
+void ExecTask::remove_instances(int n)
+{
+  Task::remove_instances(n);
+  int instance_count = (int)host_.size() - 2;
+  for (int i = instance_count - 1; i >= instance_count - n; i--)
+    host_.erase("instance_" + std::to_string(i));
+}
+
 /**
  *  @brief Default constructor.
  */
-CommTask::CommTask(const std::string& name) : Task(name) {}
+CommTask::CommTask(const std::string& name) : Task(name)
+{
+  set_load_balancing_function([]() { return "instance_0"; });
+}
 
-/** @ingroup plugin_task
+/**
  *  @brief Smart constructor.
  */
 CommTaskPtr CommTask::init(const std::string& name)
@@ -265,7 +396,7 @@ CommTaskPtr CommTask::init(const std::string& name)
   return CommTaskPtr(new CommTask(name));
 }
 
-/** @ingroup plugin_task
+/**
  *  @brief Smart constructor.
  */
 CommTaskPtr CommTask::init(const std::string& name, double bytes, Host* source, Host* destination)
@@ -273,21 +404,29 @@ CommTaskPtr CommTask::init(const std::string& name, double bytes, Host* source,
   return init(name)->set_bytes(bytes)->set_source(source)->set_destination(destination);
 }
 
-/**
- *  @brief Do one execution of the Task.
- *  @note Call the on_this_start() func.
- *  Init and start the underlying Activity.
+/** @param instance The Task instance to fire.
+ *  @note Only the dispatcher instance triggers the on_start signal.
  */
-void CommTask::fire()
+void CommTask::fire(std::string instance)
 {
-  Task::fire();
-  auto comm = Comm::sendto_init(source_, destination_)->set_name(get_name())->set_payload_size(get_amount());
-  comm->start();
-  comm->on_this_completion_cb([this](Comm const&) { complete(); });
-  store_activity(comm);
+  Task::fire(instance);
+  if (instance == "dispatcher" or instance == "collector") {
+    auto exec = Exec::init()
+                    ->set_name(get_name() + "_" + instance)
+                    ->set_flops_amount(get_amount(instance))
+                    ->set_host(instance == "dispatcher" ? source_ : destination_);
+    exec->start();
+    exec->on_this_completion_cb([this, instance](Exec const&) { complete(instance); });
+    store_activity(exec, instance);
+  } else {
+    auto comm = Comm::sendto_init(source_, destination_)->set_name(get_name())->set_payload_size(get_amount());
+    comm->start();
+    comm->on_this_completion_cb([this, instance](Comm const&) { complete(instance); });
+    store_activity(comm, instance);
+  }
 }
 
-/** @ingroup plugin_task
+/**
  *  @param source The host to set.
  *  @brief Set a new source host.
  */
@@ -297,7 +436,7 @@ CommTaskPtr CommTask::set_source(Host* source)
   return this;
 }
 
-/** @ingroup plugin_task
+/**
  *  @param destination The host to set.
  *  @brief Set a new destination host.
  */
@@ -307,7 +446,7 @@ CommTaskPtr CommTask::set_destination(Host* destination)
   return this;
 }
 
-/** @ingroup plugin_task
+/**
  *  @param bytes The amount of bytes to set.
  */
 CommTaskPtr CommTask::set_bytes(double bytes)
@@ -319,9 +458,12 @@ CommTaskPtr CommTask::set_bytes(double bytes)
 /**
  *  @brief Default constructor.
  */
-IoTask::IoTask(const std::string& name) : Task(name) {}
+IoTask::IoTask(const std::string& name) : Task(name)
+{
+  set_load_balancing_function([]() { return "instance_0"; });
+}
 
-/** @ingroup plugin_task
+/**
  *  @brief Smart Constructor.
  */
 IoTaskPtr IoTask::init(const std::string& name)
@@ -329,7 +471,7 @@ IoTaskPtr IoTask::init(const std::string& name)
   return IoTaskPtr(new IoTask(name));
 }
 
-/** @ingroup plugin_task
+/**
  *  @brief Smart Constructor.
  */
 IoTaskPtr IoTask::init(const std::string& name, double bytes, Disk* disk, Io::OpType type)
@@ -337,9 +479,8 @@ IoTaskPtr IoTask::init(const std::string& name, double bytes, Disk* disk, Io::Op
   return init(name)->set_bytes(bytes)->set_disk(disk)->set_op_type(type);
 }
 
-/** @ingroup plugin_task
+/**
  *  @param disk The disk to set.
- *  @brief Set a new disk.
  */
 IoTaskPtr IoTask::set_disk(Disk* disk)
 {
@@ -347,7 +488,7 @@ IoTaskPtr IoTask::set_disk(Disk* disk)
   return this;
 }
 
-/** @ingroup plugin_task
+/**
  *  @param bytes The amount of bytes to set.
  */
 IoTaskPtr IoTask::set_bytes(double bytes)
@@ -356,20 +497,34 @@ IoTaskPtr IoTask::set_bytes(double bytes)
   return this;
 }
 
-/** @ingroup plugin_task */
+/**
+ *  @param type The op type to set.
+ */
 IoTaskPtr IoTask::set_op_type(Io::OpType type)
 {
   kernel::actor::simcall_answered([this, type] { type_ = type; });
   return this;
 }
 
-void IoTask::fire()
+/** @param instance The Task instance to fire.
+ *  @note Only the dispatcher instance triggers the on_start signal.
+ */
+void IoTask::fire(std::string instance)
 {
-  Task::fire();
-  auto io = Io::init()->set_name(get_name())->set_size(get_amount())->set_disk(disk_)->set_op_type(type_);
-  io->start();
-  io->on_this_completion_cb([this](Io const&) { complete(); });
-  store_activity(io);
+  Task::fire(instance);
+  if (instance == "dispatcher" or instance == "collector") {
+    auto exec = Exec::init()
+                    ->set_name(get_name() + "_" + instance)
+                    ->set_flops_amount(get_amount(instance))
+                    ->set_host(disk_->get_host());
+    exec->start();
+    exec->on_this_completion_cb([this, instance](Exec const&) { complete(instance); });
+    store_activity(exec, instance);
+  } else {
+    auto io = Io::init()->set_name(get_name())->set_size(get_amount())->set_disk(disk_)->set_op_type(type_);
+    io->start();
+    io->on_this_completion_cb([this, instance](Io const&) { complete(instance); });
+    store_activity(io, instance);
+  }
 }
-
 } // namespace simgrid::s4u