Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
SMPI network calibration: tutorial
[simgrid.git] / docs / source / tuto_network_calibration / dahu_platform_ckmeans.cpp
diff --git a/docs/source/tuto_network_calibration/dahu_platform_ckmeans.cpp b/docs/source/tuto_network_calibration/dahu_platform_ckmeans.cpp
new file mode 100644 (file)
index 0000000..d645ffa
--- /dev/null
@@ -0,0 +1,141 @@
+/* Copyright (c) 2006-2021. 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 "Utils.hpp"
+#include <boost/property_tree/json_parser.hpp>
+#include <boost/property_tree/ptree.hpp>
+#include <map>
+#include <random>
+#include <simgrid/kernel/resource/NetworkModelIntf.hpp>
+#include <simgrid/s4u.hpp>
+#include <smpi/smpi.h>
+namespace sg4 = simgrid::s4u;
+
+class NormalMixture : public Sampler {
+  std::vector<std::normal_distribution<double>> mixture_;
+  std::vector<double> prob_;
+  std::mt19937& gen_;
+
+public:
+  NormalMixture(std::mt19937& gen) : gen_(gen) {}
+  void append(double mean, double stddev, double prob)
+  {
+    mixture_.push_back(std::normal_distribution<double>(mean, stddev));
+    prob_.push_back(prob);
+  }
+  double sample()
+  {
+    std::discrete_distribution<> d(prob_.begin(), prob_.end());
+    int index    = d(gen_);
+    auto& normal = mixture_[index];
+    double value = normal(gen_);
+    return value;
+  }
+};
+
+/**
+ * @brief Callback to set latency factor for a communication
+ *
+ * @param latency_base The base latency for this calibration (user-defined)
+ * @param seg Segmentation (user-defined)
+ * @param size Message size (simgrid)
+ */
+static double latency_factor_cb(double latency_base, const SegmentedRegression& seg, double size,
+                                const sg4::Host* /*src*/, const sg4::Host* /*dst*/,
+                                const std::vector<sg4::Link*>& /*links*/,
+                                const std::unordered_set<sg4::NetZone*>& /*netzones*/)
+{
+  if (size < 63305)
+    return 0.0; // no calibration for small messages
+
+  return seg.sample(size) / latency_base;
+}
+
+/**
+ * @brief Callback to set bandwidth factor for a communication
+ *
+ * @param bw_base The base bandwidth for this calibration (user-defined)
+ * @param seg Segmentation (user-defined)
+ * @param size Message size (simgrid)
+ */
+static double bw_factor_cb(double bw_base, const SegmentedRegression& seg, double size, const sg4::Host* /*src*/,
+                           const sg4::Host* /*dst*/, const std::vector<sg4::Link*>& /*links*/,
+                           const std::unordered_set<sg4::NetZone*>& /*netzones*/)
+{
+  if (size < 63305)
+    return 1.0; // no calibration for small messages
+  double est_bw = 1.0 / seg.get_coef(size);
+  return est_bw / bw_base;
+}
+
+static double smpi_cost_cb(const SegmentedRegression& seg, double size, sg4::Host* /*src*/, sg4::Host* /*dst*/)
+{
+  return seg.sample(size);
+}
+
+static SegmentedRegression read_json_file(const std::string& jsonFile, std::mt19937& gen, bool read_coef = true)
+{
+  boost::property_tree::ptree pt;
+  boost::property_tree::read_json(jsonFile, pt);
+  std::unordered_map<double, std::shared_ptr<NormalMixture>> mixtures;
+  std::unordered_map<double, double> coefs;
+  printf("Starting parsing file: %s\n", jsonFile.c_str());
+  pt = pt.get_child("seg"); // go to segments part
+  SegmentedRegression seg(read_coef);
+  for (const auto& it : pt) {
+    double max    = it.second.get_child("max_x").get_value<double>();
+    coefs[max]    = it.second.get_child("coefficient").get_value<double>();
+    auto& mixture = mixtures[max];
+    if (!mixture)
+      mixture = std::make_shared<NormalMixture>(gen);
+    mixture->append(it.second.get_child("mean").get_value<double>(), it.second.get_child("sd").get_value<double>(),
+                    it.second.get_child("prob").get_value<double>());
+  }
+  for (const auto& i : mixtures) {
+    seg.append(i.first, coefs[i.first], i.second);
+  }
+  return seg;
+}
+
+/** @brief Programmatic version of dahu */
+extern "C" void load_platform(const sg4::Engine& e);
+void load_platform(const sg4::Engine& e)
+{
+  // setting threshold sync/async modes
+  e.set_config("smpi/async-small-thresh", 63305);
+  e.set_config("smpi/send-is-detached-thresh", 63305);
+
+  /* reading bandwidth and latency base value. It is the same for all regressions, get from first file */
+  boost::property_tree::ptree pt;
+  boost::property_tree::read_json("pingpong_ckmeans.json", pt);
+  double bw_base  = pt.get_child("bandwidth_base").get_value<double>();
+  double lat_base = pt.get_child("latency_base").get_value<double>();
+  printf("Read bandwidth_base: %e latency_base: %e\n", bw_base, lat_base);
+
+  load_dahu_platform(e, bw_base, lat_base);
+
+  static std::mt19937 gen(42); // remove it from stack, since we need it after this this load_platform function is over
+
+  /* setting network factors callbacks */
+  simgrid::kernel::resource::NetworkModelIntf* model = e.get_netzone_root()->get_network_model();
+
+  SegmentedRegression seg = read_json_file("pingpong_ckmeans.json", gen, false);
+  model->set_lat_factor_cb(std::bind(&latency_factor_cb, lat_base, seg, std::placeholders::_1, std::placeholders::_2,
+                                     std::placeholders::_3, std::placeholders::_4, std::placeholders::_5));
+
+  model->set_bw_factor_cb(std::bind(&bw_factor_cb, bw_base, seg, std::placeholders::_1, std::placeholders::_2,
+                                    std::placeholders::_3, std::placeholders::_4, std::placeholders::_5));
+
+  seg = read_json_file("send_ckmeans.json", gen);
+  smpi_register_op_cost_callback(SmpiOperation::SEND, std::bind(&smpi_cost_cb, seg, std::placeholders::_1,
+                                                                std::placeholders::_2, std::placeholders::_3));
+
+  seg = read_json_file("isend_ckmeans.json", gen);
+  smpi_register_op_cost_callback(SmpiOperation::ISEND, std::bind(&smpi_cost_cb, seg, std::placeholders::_1,
+                                                                 std::placeholders::_2, std::placeholders::_3));
+  seg = read_json_file("recv_ckmeans.json", gen);
+  smpi_register_op_cost_callback(SmpiOperation::RECV, std::bind(&smpi_cost_cb, seg, std::placeholders::_1,
+                                                                std::placeholders::_2, std::placeholders::_3));
+}
\ No newline at end of file