XGC1
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Pages
load_balance.hpp
Go to the documentation of this file.
1 #ifndef LOAD_BALANCE_HPP
2 #define LOAD_BALANCE_HPP
3 
4 #include "timer_macro.hpp"
5 #include "shift.hpp"
6 #include "count_ptl_per_node.hpp"
7 #include "view_arithmetic.hpp"
8 #include "f0_redistribute.hpp"
9 
10 // Used for original fortran load balancer
11 extern "C" void calculate_load_imbalance(double f0_cost);
12 extern "C" int assess_whether_to_rebalance_load();
13 extern "C" void reset_cost_trackers();
14 extern "C" void set_weights(int* gvid0_pid, double* ptl_count, double* f0_node_cost);
15 
16 class LoadRegion{
17  public:
18  enum class UpdateMethod{
19  NoHistory=0,
21  };
22 
23  private:
24 
25  View<double*,HostType> estimated_time_per_vertex;
26  std::string region_name;
27  std::vector<std::string> timer_names;
28  bool verbose;
35 
36 #ifdef USE_MPI
37  MPI_Comm inter_period_comm;
38  MPI_Comm period_comm;
39 #endif
40  int n_periods;
43 
45 
47 
48  void update_model_no_history(const View<int*,CLayout,HostType>& current_partition, const View<double*, HostType>& all_periods_timings){
49  // Loop over ranks' timers
50  for(int i=0; i<all_periods_timings.size(); i++){
51  int node_offset = current_partition(i) - 1;
52  int next_node_offset = current_partition(i+1) - 1;
53  int nnodes = next_node_offset - node_offset;
54  // Best 0th order estimate is to assume all vertices took the same time
55  double time_per_node = all_periods_timings(i)/nnodes;
56  for(int j=0; j<nnodes; j++){
57  estimated_time_per_vertex(j + node_offset) = time_per_node;
58  }
59  }
60  }
61 
62  void update_model_exp_history(const View<int*,CLayout,HostType>& current_partition, const View<double*, HostType>& all_periods_timings){
63  // Loop over ranks' timers
64  for(int i=0; i<all_periods_timings.size(); i++){
65  // Get partition owned by this rank
66  int node_offset = current_partition(i) - 1;
67  int next_node_offset = current_partition(i+1) - 1;
68  int nnodes = next_node_offset - node_offset;
69 
70  // Get average time per node for this rank
71  double avg_time_per_node = all_periods_timings(i)/nnodes;
72 
73  // Get expected time based on existing model
74  double expected_time = 0.0;
75  for(int j=0; j<nnodes; j++){
76  expected_time += estimated_time_per_vertex(j + node_offset);
77  }
78  double expected_time_per_node = expected_time/nnodes;
79 
80  // Difference between observed time and expected time
81  double extra_time_per_node = avg_time_per_node - expected_time_per_node;
82 
83  for(int j=0; j<nnodes; j++){
84  // Distribute extra time evenly
85  constexpr double adjustment_rate = 0.5;
86 
87  estimated_time_per_vertex(j + node_offset) += extra_time_per_node*adjustment_rate;
88  }
89  }
90  }
91 
93  // Get time from camtimers (accumulated)
94  double new_accumulated_time = 0.0;
95 
96  // Loop through all timers associated with this region, and add the accumulated time for each
97  for(int i=0; i<timer_names.size(); i++){
98  double accumulated_single_timer = 0;
99  int ierr = GPTLget_wallclock(timer_names[i].c_str(), -1, &accumulated_single_timer);
100 
101  // If ierr != 0, timer is not yet defined, so the accumulated time is zero
102  if(ierr != 0) accumulated_single_timer = 0.0;
103 
104  new_accumulated_time += accumulated_single_timer;
105  }
106 
107  // Subtract previous accumulated time to get time spent since last update
108  double time_spent = new_accumulated_time - time_accumulated;
109 
110  // Save new accumulated time
111  time_accumulated = new_accumulated_time;
112 
113  return time_spent;
114  }
115 
116  void touch_timers(){
117  for(int i=0; i<timer_names.size(); i++){
118  GPTLstart(timer_names[i].c_str());
119  GPTLstop(timer_names[i].c_str());
120  }
121  }
122 
123  public:
124 
125 #ifdef USE_MPI
126  LoadRegion(int n_vertices, const MPI_Comm& inter_period_comm, const MPI_Comm& period_comm, UpdateMethod update_method, bool verbose, std::string region_name, std::vector<std::string> timer_names)
127  : model_is_initialized(false),
128  verbose(verbose),
129  time_accumulated(0.0),
130  inter_period_comm(inter_period_comm),
131  period_comm(period_comm),
132  estimated_time_per_vertex("estimated_time_per_vertex",n_vertices),
134  update_method(update_method),
135  model_has_history(false),
136  region_name(region_name),
137  timer_names(timer_names)
138  {
139  // Touch GPTL timers to be sure they exist
140  touch_timers();
141 
142  // Initialize timer
143  reset_timer();
144 
145  // Get problem size from comms
146  MPI_Comm_rank(period_comm, &my_period_rank);
147  MPI_Comm_size(period_comm, &n_unique_ranks);
148  MPI_Comm_size(inter_period_comm, &n_periods);
149  }
150 #endif
151 
152  void reset_timer(){
153  // Reset timer by getting the time since the previous call and suppressing the output
154  double time_spent = get_time_since_previous_call();
155  }
156 
157  double get_estimated_time_per_vertex(int i) const{
158  return estimated_time_per_vertex(i);
159  }
160 
161  View<double*,HostType> get_estimated_time_per_vertex() const{
163  }
164 
166  return prediction_undershoot;
167  }
168 
171  }
172 
175  }
176 
178  return model_is_initialized;
179  }
180 
181  // Predict performance of new partition based on model
182  double get_largest_predicted_time(const View<int*,CLayout,HostType>& proposed_partition) const{
183  double largest_time = 0.0;
184  //int largest_time_ind = 0;
185  for(int i=0; i<(proposed_partition.size()-1); i++){
186  double proc_time = 0.0;
187  for(int i_node=proposed_partition(i)-1; i_node<proposed_partition(i+1)-1; i_node++){
188  proc_time += estimated_time_per_vertex(i_node);
189  }
190  if(proc_time>largest_time){
191  //largest_time_ind = i;
192  largest_time = proc_time;
193  }
194  }
195  return largest_time;
196  }
197 
199  reset_timer();
200 
201  // Model can now be used
202  model_is_initialized = true;
203  }
204 
205  void update_model(const View<int*,CLayout,HostType>& current_partition, double manual_time=-1.0){
207  // Get time spent since last model update
208  // Time can be entered manually; it's used if positive (i.e. the input was provided)
209  double time_spent_this_rank = (manual_time<0.0 ? get_time_since_previous_call()
210  : manual_time);
211 
212  // Reduce onto plane 0
213  // Could try MPI_SUM rather than MPI_MAX
214  double time_spent;
215  if(n_periods>1){
216 #ifdef USE_MPI
217  MPI_Reduce(&time_spent_this_rank, &time_spent, 1, MPI_DOUBLE, MPI_MAX, 0, inter_period_comm);
218 #endif
219  }else{
220  time_spent = time_spent_this_rank;
221  }
222 
223  // Allocate for timing of each
224  View<double*, HostType> all_periods_timings(NoInit("all_periods_timings"), n_unique_ranks);
225 #ifdef USE_MPI
226  // Gather from all ranks in plane 0
227  MPI_Gather(&time_spent, 1, MPI_DOUBLE, all_periods_timings.data(), 1, MPI_DOUBLE, 0, period_comm);
228 #endif
229 
230  if (is_rank_zero()){
231  // Look at the timing prediction made last time
233  if(verbose) printf("\nPredicted max time in this region was %1.5e", predicted_max_region_time);
234 
235  // Get max region time
237  double observed_sum_region_time = 0.0;
238  for(int i=0; i<all_periods_timings.size(); i++){
239  observed_max_region_time = std::max(observed_max_region_time, all_periods_timings(i));
240  observed_sum_region_time += all_periods_timings(i);
241  }
242  double observed_idle_region_time = observed_max_region_time*all_periods_timings.size() - observed_sum_region_time;
243  observed_load_imbalance = observed_idle_region_time/observed_sum_region_time;
244  if(verbose) printf("\nObserved max time in this region was %1.5e", observed_max_region_time);
245  if(verbose) printf("\n - Load imbalance (T_idle/T_work): %1.2f%%", observed_load_imbalance*100);
246 
247  // Take the max here so that there is never an assumed overshoot
248  if(predicted_max_region_time!=0.0){ // If time is zero, there hasn't been a prediction yet
250  }
251  if(verbose) printf("\nThe prediction undershot by a factor of %1.5e", observed_max_region_time/predicted_max_region_time);
252 
254  update_model_no_history(current_partition, all_periods_timings);
256  if(model_has_history){
257  update_model_exp_history(current_partition, all_periods_timings);
258  }else{
259  update_model_no_history(current_partition, all_periods_timings);
260  model_has_history = true;
261  }
262  }
263  }
264  }else{
265  exit_XGC("Error: Update method not available\n");
266  }
267  }
268 };
269 
271  public:
272  enum class WeightingAlgorithm{
275  Fortran,
276  Default
277  };
278 
279  enum class ConstraintOption{
280  ParticleCount=0
281  };
282 
283  enum class ReweightOption{
284  Always=0,
285  IfBetter
286  };
287 
288  private:
289 
291 
292  std::vector<LoadRegion> regions;
293 
294  View<int*,CLayout,HostType> proposed_partition;
295 
297  bool verbose;
298 
300 
301  double get_even_division(const View<double*,HostType>& input, int n) const{
302  double total = 0.0;
303  Kokkos::parallel_reduce("sum", Kokkos::RangePolicy<HostExSpace>(0,input.size()), [=]( const int i, double& l_total){
304  l_total += input(i);
305  }, total);
306 
307  return (total/n);
308  }
309 
310 
311  bool greedily_fill_partition(const View<double*,HostType>& weight, const View<double*,HostType>& constraint1, double target_weight_per_rank){
312  int nnode = weight.size();
313  int nproc = proposed_partition.size()-1;
314 
315  // Start from rank 0 of the plane; assign nodes until it meets the desired weight per rank, then move to next rank
316  double constraint1_in_this_rank = 0.0;
317  double weight_in_this_rank = 0.0;
318  proposed_partition(0) = 1; // 1-indexed
319  int pid = 0;
320  bool assign_one_node_per_proc = false;
321  for(int i_node = 0; i_node<nnode; i_node++){
322  // Since every rank needs at last one node, switch to assigning at every
323  // iteration if there are only as many nodes left as ranks
324  if(nnode-i_node==nproc-pid) assign_one_node_per_proc = true;
325 
326  // Check if a criterion is met to consider the rank sufficiently loaded
327  bool rank_is_loaded = false;
328 
329  // Criterion 1: We are out of nodes
330  if(assign_one_node_per_proc) rank_is_loaded = true;
331 
332  // Criterion 2: We have hit a constraint
333  if(constraint1_in_this_rank>=constraint1_max) rank_is_loaded = true;
334 
335  // Criterion 3: We have loaded the rank with an even amount of work
336  if(weight_in_this_rank>=target_weight_per_rank) rank_is_loaded = true;
337 
338  // If there are enough particles assigned to this rank, move to the next rank
339  if(rank_is_loaded){
340  proposed_partition(pid+1) = i_node+1; // 1-indexed
341  if(verbose) printf("\nRank %d is loaded (Nodes %d-%d); weight=%1.4e, ptl*nplanes=%d", pid, proposed_partition(pid), proposed_partition(pid+1)-1, weight_in_this_rank, int(constraint1_in_this_rank));
342  constraint1_in_this_rank = 0.0;
343  weight_in_this_rank = 0.0;
344  pid++;
345  if(pid==nproc-1) break;
346  }
347  constraint1_in_this_rank += constraint1(i_node);
348  weight_in_this_rank += weight(i_node);
349  }
350 
351  // Last value will always be nnode+1
352  proposed_partition(nproc) = nnode+1;
353 
354  // Check n_ptl in final rank
355  constraint1_in_this_rank = 0.0;
356  for(int i_node=proposed_partition(nproc-1)-1; i_node<nnode; i_node++){
357  constraint1_in_this_rank += constraint1(i_node);
358  }
359  if(constraint1_in_this_rank>=constraint1_max){
360  // Return false if the constraint was not met
361  return false;
362  }else{
363  return true;
364  }
365  }
366 
367  // Predict performance of new partition based on model
368  double get_largest_predicted_time(const View<int*,CLayout,HostType>& partition, const View<double*,HostType>& weight) const{
369  double largest_time = 0.0;
370  //int largest_time_ind = 0;
371  for(int i=0; i<(partition.size()-1); i++){
372  double proc_time = 0.0;
373  for(int i_node=partition(i)-1; i_node<partition(i+1)-1; i_node++){
374  proc_time += weight(i_node);
375  }
376  if(proc_time>largest_time){
377  //largest_time_ind = i;
378  largest_time = proc_time;
379  }
380  }
381  return largest_time;
382  }
383 
384  void one_weight_balance(const View<double*,HostType>& weight, const View<double*,CLayout,HostType> constraint1){
385  int nproc = proposed_partition.size()-1;
386 
387  // Ideally, each rank would get the same amount of work, i.e. the average
388  double ideal_weight_per_rank = get_even_division(weight, nproc);
389 
390  // Make initial attempt, targeting even distribution of work
391  bool meets_constraints = greedily_fill_partition(weight, constraint1, ideal_weight_per_rank);
392 
393  if(!meets_constraints){
394  // An equal work distribution cannot be assigned due to a constraint.
395  // First, determine the weights if only the constraint were followed
396  meets_constraints = greedily_fill_partition(constraint1, constraint1, get_even_division(constraint1, nproc));
397  if(!meets_constraints) exit_XGC("\nUnexpected issue in load balance: constraint-based partition doesn't satisfy constraints\n");
398 
399  // This partition meets the constraints, but the proposed timing is likely unacceptable
400  double upper_limit_weight_per_rank = get_largest_predicted_time(proposed_partition, weight);
401  if(verbose) printf("The ideal distribution (%1.3e) does not satisfy constraints. This distribution (%1.3e) does.\n", ideal_weight_per_rank, upper_limit_weight_per_rank);
402 
403  // Bisect the difference between the original target time and the upper limit, and see if that satisfies the constraints
404  // Continue until the load balance has been minimized to a precision based on the the original desired time
405  const double desired_precision = 0.01; // Fraction of imbalance that's worth even checking whether it can be removed
406  double desired_step_size = ideal_weight_per_rank*desired_precision;
407  double step_size = upper_limit_weight_per_rank - ideal_weight_per_rank;
408  double compromise_weight_per_rank = upper_limit_weight_per_rank;
409  if(verbose) printf("\nEmploying a binary search to narrow down on the best option.");
410  while(step_size>desired_step_size){
411  // Halve the size of the step
412  step_size /= 2;
413 
414  if(meets_constraints){
415  // Try reducing the load imbalance since constraints were met
416  compromise_weight_per_rank -= step_size;
417  }else{
418  // Try raising the load imbalance since constraints were broken
419  compromise_weight_per_rank += step_size;
420  }
421 
422  // Get new partition
423  meets_constraints = greedily_fill_partition(weight, constraint1, compromise_weight_per_rank);
424  if(verbose) printf("\n Stepped by %1.3e to %1.3e. The new partition does%s meet constraints.", step_size, compromise_weight_per_rank, meets_constraints?"" : "nt");
425  }
426  // In case we end at a partition that fails
427  while(!meets_constraints){
428  compromise_weight_per_rank += step_size;
429  meets_constraints = greedily_fill_partition(weight, constraint1, compromise_weight_per_rank);
430  if(verbose) printf("\n Stepped by %1.3e UP to %1.3e. The new partition does%s meet constraints.", step_size, compromise_weight_per_rank, meets_constraints?"" : "nt");
431  }
432  if(verbose) printf("\n");
433  }
434  }
435 
436  // TODO
437  // Evaluate overall model performance
438  // MAIN_LOOP time vs sum of regions (% coverage)
439  //
440 
441  /* This function assesses the proposed partition and decides whether it is worth switching to.
442  * The simplest formulation is to recommend the new partition if the new partition is predicted to be faster
443  * than the existing one. There is also a factor to account for historical inaccuracy of the previous estimate.
444  * A future option would be to require a certain level of improvement before repartitioning. */
446  // Total time spent in regions (assuming that global barriers demarcate them):
447  double observed_max_all_rgns_time = 0.0;
448  for(int i=0; i<regions.size(); i++){
449  observed_max_all_rgns_time += regions[i].get_observed_max_region_time();
450  }
451 
452  // Total predicted time, accounting for observed prediction undershoot for each region
453  double predicted_max_all_rgns_time = 0.0;
454  for(int i=0; i<regions.size(); i++){
455  predicted_max_all_rgns_time += regions[i].get_largest_predicted_time(proposed_partition)*regions[i].get_prediction_undershoot();
456  }
457 
458  double fractional_improvement = 1.0 - predicted_max_all_rgns_time/observed_max_all_rgns_time;
459 
460  if(verbose){
461  printf("\nDetermining whether to adopt the proposed partition:");
462  printf("\n Observed time with current partition was: %1.3e", observed_max_all_rgns_time);
463  printf("\n Predicted time with proposed partition, adjusting for historical undershoot (of Rgn 0: %1.3e): %1.3e", regions[0].get_prediction_undershoot(), predicted_max_all_rgns_time);
464  printf("\n Adopted if Fractional improvement of adjusted prediction (%1.3e) exceeds specified threshold (%1.3e)\n", fractional_improvement, threshold_to_rebalance);
465  }
466 
467  // Recommend the new partition if the predicted time is less than the time observed with the current partition
468  if(fractional_improvement > threshold_to_rebalance){
469  return true;
470  }else{
471  return false;
472  }
473  }
474 
476  // Print new partition
477  for(int i = 0; i<proposed_partition.size()-1; i++){
478  int nnodes_for_proc = proposed_partition(i+1) - proposed_partition(i);
479  printf("process %d : nnodes = %d; range = (%d - %d)\n", i, nnodes_for_proc, proposed_partition(i), proposed_partition(i+1)-1);
480  }
481  }
482 
483  void update_model(const View<int*,CLayout,HostType>& current_partition){
484  for(int i=0; i<regions.size(); i++){
485  regions[i].update_model(current_partition);
486  }
487  }
488 
489  // Overload if providing timings manually rather than using camtimers. This is used for testing.
490  void update_model(const View<int*,CLayout,HostType>& current_partition, const std::vector<double>& manual_times){
491  for(int i=0; i<regions.size(); i++){
492  regions[i].update_model(current_partition, manual_times[i]);
493  }
494  }
495 
497  for(int i=0; i<regions.size(); i++){
498  regions[i].initialize_model();
499  }
500  }
501 
503  bool is_initialized=true;
504  for(int i=0; i<regions.size(); i++){
505  is_initialized = (is_initialized && regions[i].get_model_is_initialized());
506  }
507  return is_initialized;
508  }
509 
510  void propose_new_partition(const Kokkos::View<double*,Kokkos::LayoutRight,HostType>& ptl_count, WeightingAlgorithm weighting_algorithm){
511  auto constraint1 = ptl_count; // Constrain based on particle count (memory)
512  if(weighting_algorithm == WeightingAlgorithm::SingleRegionBalance){
513  if(regions.size()!=1) exit_XGC("Error: Load balancing is currently single-region only\n");
514  one_weight_balance(regions[0].get_estimated_time_per_vertex(), constraint1);
515  }else if(weighting_algorithm == WeightingAlgorithm::ParticleBalance){
516  one_weight_balance(ptl_count, constraint1);
517  }
518 
519  if(verbose){
520  printf("\nNewly PROPOSED partition:\n");
522  }
523  }
524 
526  const VelocityGrid& vgrid, Plasma& plasma, DomainDecomposition<DeviceType>& pol_decomp, WeightingAlgorithm weighting_algorithm){
527  if(weighting_algorithm==WeightingAlgorithm::Fortran){
528  // In the fortran algorithm the new distribution is calculated and set here
529  // Count number of particles belonging to each node
530  Kokkos::View<double**,Kokkos::LayoutRight,HostType> ptl_count = count_ptl_per_node_elec_main_ion(grid, magnetic_field, plasma, sml.electron_on);
531 
532  if(pol_decomp.gvid0_pid_h.size()==2){
533  // Skip set_weights if there is only one rank per plane
534  pol_decomp.gvid0_pid_h(0) = 1;
535  pol_decomp.gvid0_pid_h(1) = pol_decomp.nnodes_on_plane + 1;
536  }else{
537  TIMER("SET_WEIGHTS",
538  set_weights(pol_decomp.gvid0_pid_h.data(), ptl_count.data(), plasma.f0_node_cost.data()) );
539  }
540  }else{
541 #ifdef USE_MPI
542  // Broadcast proposal to all ranks
543  MPI_Bcast(proposed_partition.data(), proposed_partition.size(), MPI_INT, 0, SML_COMM_WORLD);
544 #endif
545 
546  // Copy proposal to pol_decomp
547  Kokkos::deep_copy(pol_decomp.gvid0_pid_h, proposed_partition);
548 
549  if(is_rank_zero()){
550  printf("\nNEW PARTITION:\n");
552  }
553  }
554 
555  // Update bounds and device copy of the decomposition array
556  pol_decomp.update_pol_decomp();
557  }
558 
560  const VelocityGrid& vgrid, Plasma& plasma, DomainDecomposition<DeviceType>& pol_decomp,
561  const View<int*,CLayout,HostType>& old_partition){
562  if (sml.f0_grid){
563  // Move f information to correct MPI rank after decomposition update
564  TIMER("F0_REDISTRIBUTE",
565  f0_redistribute(plasma, pol_decomp, grid, magnetic_field, vgrid, old_partition) );
566  }
567 
568  // Shift particles to correct MPI rank after decomposition update
569  TIMER("SHIFT_R",
570  shift_all_species(plasma, grid, magnetic_field, pol_decomp, Shift::NoPhase0) );
571  }
572 
573  bool will_rebalance(ReweightOption reweight_option, WeightingAlgorithm weighting_algorithm, double f0_cost){
574  if(reweight_option==ReweightOption::Always){
575  // No need to check if model is an improvement
576  return true;
577  }else{
578  if(weighting_algorithm==WeightingAlgorithm::Fortran){
579  // Calculate load imbalance
580  calculate_load_imbalance(f0_cost);
581 
582  // Assess if load imbalance merits a rebalancing
583  int should_rebalance_int = assess_whether_to_rebalance_load();
584 
585  // Reset fortran cost trackers
587 
588  return (should_rebalance_int==1);
589  }else{
590  // Evaluate the proposed partition on rank 0
591  bool should_rebalance;
592  if(weighting_algorithm==WeightingAlgorithm::ParticleBalance){
593  // Since ParticleBalance doesn't contain an internal model, cant evaluate the partition
594  // so just always approve, for now.
595  // Probably better to have the model updatable by something other than time so it can still
596  // be assessed in this mode
597  return true;
598  }else{
599  if(is_rank_zero()) should_rebalance = recommend_proposed_partition();
600  // Convert to int because MPI_C_BOOL documentation is confusing
601  int should_rebalance_int = (should_rebalance ? 1 : 0);
602 #ifdef USE_MPI
603  MPI_Bcast(&should_rebalance_int, 1, MPI_INT, 0, SML_COMM_WORLD);
604 #endif
605  return (should_rebalance_int==1);
606  }
607  }
608  }
609  }
610 
611  public:
612 
614  : proposed_partition(NoInit("proposed_partition"), pol_decomp.gvid0_pid_h.layout())
615  {
616  double max_mem_redist_gb = 10.0;
617  std::string weighting_algorithm_str = "Fortran";
618  std::string update_method_str = "NoHistory";
619  if(nlr.namelist_present("load_balance_param")){
620  nlr.use_namelist("load_balance_param");
621  max_mem_redist_gb = nlr.get<double>("max_mem_redist_gb", 10.0);
622  weighting_algorithm_str = nlr.get<std::string>("weighting_algorithm", "Fortran");
623  threshold_to_rebalance = nlr.get<double>("threshold_to_rebalance", 0.02);
624  verbose = nlr.get<bool>("verbose", false);
625  update_method_str = nlr.get<std::string>("update_method", "NoHistory");
626  }else{
627  threshold_to_rebalance = 0.02;
628  verbose = false;
629  }
630 
631  // Simple model for memory usage from load balance - just particles
632  double max_n_ptl_on_rank = max_mem_redist_gb*1024*1024*1024/80.0; // ~ 80 B per particle
633 
634  if(weighting_algorithm_str=="Fortran"){
636  }else if(weighting_algorithm_str=="ParticleBalance"){
638  }else if(weighting_algorithm_str=="SingleRegionBalance"){
640  }else{
641  exit_XGC("\nError: weighting_algorithm input not valid.");
642  }
643 
644 #ifdef USE_MPI
645  // Set up load balance regions
647  LoadRegion::UpdateMethod update_method;
648  if(update_method_str=="NoHistory"){
649  update_method = LoadRegion::UpdateMethod::NoHistory;
650  }else if(update_method_str=="ExpHistory"){
651  update_method = LoadRegion::UpdateMethod::ExpHistory;
652  }else{
653  exit_XGC("\nError: update_method input not valid.");
654  }
655 
656  regions.push_back(LoadRegion(pol_decomp.nnodes_on_plane, pol_decomp.mpi.intpl_comm, pol_decomp.mpi.plane_comm, update_method, verbose,
657  "Push",
658  {"ipc1:PUSHE",
659  "ipc1:PUSHI",
660  "ipc2:PUSHE",
661  "ipc2:PUSHI"}));
662  }
663 
664  // Hard-code for now
666 
667  if(constraint_option == ConstraintOption::ParticleCount){
668  // Normalize by n_planes since we are using the sum of planes as the constraint
669  constraint1_max = max_n_ptl_on_rank*pol_decomp.mpi.n_intpl_ranks;
670  }else{
671  exit_XGC("\nError: ParticleConstraint is only available ConstraintOption.\n");
672  }
673 #endif
674  }
675 
677  const VelocityGrid& vgrid, Plasma& plasma, DomainDecomposition<DeviceType>& pol_decomp, ReweightOption reweight_option,
678  WeightingAlgorithm weighting_algorithm = WeightingAlgorithm::Default){
679 
680  // Skip rebalancing if poloidal decomposition is turned off. Normally we should put the function in
681  // an if statement rather than using return; kept here in case we might want some balancing for some reason
682  // even without poloidal decomposition
683  if(pol_decomp.pol_decomp==false) return;
684 
685  // Rebalance is trivial if there is only one rank on a plane (generalize this from no-MPI case)
686 #ifdef USE_MPI
687 
688  if(weighting_algorithm == WeightingAlgorithm::Default){
689  weighting_algorithm = default_weighting_algorithm;
690  }
691 
692  if(default_weighting_algorithm==WeightingAlgorithm::Fortran){
693  // Override input algorithm; if the default is Fortran, it should always use the fortran load balancer
694  weighting_algorithm = default_weighting_algorithm;
695  }
696 
697  if(is_rank_zero() && verbose){
698  if(weighting_algorithm==WeightingAlgorithm::Fortran) printf("\nLoad balance called with Fortran algorithm\n");
699  else if(weighting_algorithm==WeightingAlgorithm::ParticleBalance) printf("\nLoad balance called with Ptl algorithm\n");
700  else if(weighting_algorithm==WeightingAlgorithm::SingleRegionBalance) printf("\nLoad balance called with SingleRegion algorithm\n");
701  }
702 
703  GPTLstart("REBALANCE");
704  if(weighting_algorithm!=WeightingAlgorithm::Fortran){
705  GPTLstart("PTL_COUNT");
707  long long int total_n_ptl = species.get_total_n_ptl();
708  int max_n_ptl = species.get_max_n_ptl();
709  double avg_n_ptl = (double)(total_n_ptl)/pol_decomp.mpi.nranks;
710  double max_ratio = max_n_ptl/avg_n_ptl;
711  if(is_rank_zero()) printf(" Species %d: (max/avg ptl per rank = %1.2f); total n_ptl = %lld\n", species.idx, max_ratio, total_n_ptl);
713  GPTLstop("PTL_COUNT");
714 
715  GPTLstart("PTL_COUNT_PER_NODE");
716  auto ptl_count = count_all_ptl_per_node(grid, magnetic_field, plasma);
717  GPTLstop("PTL_COUNT_PER_NODE");
718 
719  if(weighting_algorithm!=WeightingAlgorithm::ParticleBalance){
720  if(model_is_initialized()){
721  // Update the model with the latest timing
722  TIMER("LOAD_BAL_UPDATE",
723  update_model(pol_decomp.gvid0_pid_h) );
724  }else{
725  if(is_rank_zero() && verbose) printf("Initializing timing model, no partition proposal yet.");
726  initialize_model();
727  GPTLstop("REBALANCE");
728  return;
729  }
730  }
731 
732  // With the updated model, proposed a new partition
733  GPTLstart("LOAD_BAL_NEW_PART");
734  if (is_rank_zero()) propose_new_partition(ptl_count, weighting_algorithm);
735  GPTLstop("LOAD_BAL_NEW_PART");
736  }
737 
738  GPTLstart("LOAD_BAL_REBAL");
739  double f0_cost = sml.f0_grid ? sum_view(plasma.f0_node_cost) : 0.0;
740  if(will_rebalance(reweight_option, weighting_algorithm, f0_cost)){
741  // Save the old partition, it is used to save some communication in the f0 redistribution
742  View<int*,CLayout,HostType> old_partition(NoInit("old_partition"), pol_decomp.gvid0_pid_h.layout());
743  Kokkos::deep_copy(old_partition, pol_decomp.gvid0_pid_h);
744 
745  // Update pol_decomp
746  TIMER("LOAD_BAL_SET_NEW",
747  set_new_partition(sml, grid, magnetic_field, vgrid, plasma, pol_decomp, weighting_algorithm) );
748 
749  // Update particles and f0
750  TIMER("LOAD_BAL_REDIST",
751  redistribute_load(sml, grid, magnetic_field, vgrid, plasma, pol_decomp, old_partition) );
752  }
753  GPTLstop("LOAD_BAL_REBAL");
754  GPTLstop("REBALANCE");
755 #endif
756  }
757 
758 
759  // More generic rebalance
760  void rebalance(DomainDecomposition<DeviceType>& pol_decomp, const View<double*,CLayout,HostType>& constraint, const std::vector<double>& timings, double& load_imbalance, View<double*,HostType>& model_belief){
761 #ifdef USE_MPI
762  update_model(pol_decomp.gvid0_pid_h, timings);
763 
764  load_imbalance = regions[0].get_observed_load_imbalance();
765  model_belief = regions[0].get_estimated_time_per_vertex();
766 
767  // Propose new partition
768  if (is_rank_zero()) propose_new_partition(constraint, default_weighting_algorithm);
769 
770  double f0_cost = 0.0;
771  if(will_rebalance(ReweightOption::IfBetter, default_weighting_algorithm, f0_cost)){
772  // Broadcast proposal to all ranks
773  MPI_Bcast(proposed_partition.data(), proposed_partition.size(), MPI_INT, 0, SML_COMM_WORLD);
774 
775  // Copy proposal to pol_decomp
776  Kokkos::deep_copy(pol_decomp.gvid0_pid_h, proposed_partition);
777 
778  // Update bounds and device copy of the decomposition array
779  pol_decomp.update_pol_decomp();
780 
781  // If there were a real load, redistribute it here
782  }
783 #endif
784  }
785 };
786 
787 #endif
bool model_has_history
Definition: load_balance.hpp:30
View< double *, HostType > estimated_time_per_vertex
Definition: load_balance.hpp:25
static int GPTLstart(const char *name)
Definition: timer_macro.hpp:9
bool will_rebalance(ReweightOption reweight_option, WeightingAlgorithm weighting_algorithm, double f0_cost)
Definition: load_balance.hpp:573
bool is_rank_zero()
Definition: globals.hpp:27
void one_weight_balance(const View< double *, HostType > &weight, const View< double *, CLayout, HostType > constraint1)
Definition: load_balance.hpp:384
MPI_Comm SML_COMM_WORLD
Definition: my_mpi.cpp:4
double threshold_to_rebalance
Definition: load_balance.hpp:296
T get(const string &param, const T default_val, int val_ind=0)
Definition: NamelistReader.hpp:373
double time_accumulated
Definition: load_balance.hpp:44
double predicted_max_region_time
Definition: load_balance.hpp:31
double observed_max_region_time
Definition: load_balance.hpp:32
double get_largest_predicted_time(const View< int *, CLayout, HostType > &partition, const View< double *, HostType > &weight) const
Definition: load_balance.hpp:368
Definition: velocity_grid.hpp:8
Definition: sml.hpp:8
void update_model_exp_history(const View< int *, CLayout, HostType > &current_partition, const View< double *, HostType > &all_periods_timings)
Definition: load_balance.hpp:62
int n_unique_ranks
How many ranks are in a &#39;period&#39; (in tokamaks, in a plane)
Definition: load_balance.hpp:41
subroutine plasma(grid, itr, p, dene_out, deni_out, Te_out, Ti_out, Vparai_out)
Calculate the plasma density, temperature, and parallel velocity for a point in triangle itr using pl...
Definition: neutral_totalf.F90:1237
void rebalance(DomainDecomposition< DeviceType > &pol_decomp, const View< double *, CLayout, HostType > &constraint, const std::vector< double > &timings, double &load_imbalance, View< double *, HostType > &model_belief)
Definition: load_balance.hpp:760
Definition: plasma.hpp:109
Definition: NamelistReader.hpp:193
Definition: magnetic_field.hpp:12
void propose_new_partition(const Kokkos::View< double *, Kokkos::LayoutRight, HostType > &ptl_count, WeightingAlgorithm weighting_algorithm)
Definition: load_balance.hpp:510
int idx
Index in all_species.
Definition: species.hpp:78
bool verbose
Definition: load_balance.hpp:297
void for_all_nonadiabatic_species(F func, DevicePtlOpt device_ptl_opt=UseDevicePtl)
Definition: plasma.hpp:128
double prediction_undershoot
Definition: load_balance.hpp:34
double get_even_division(const View< double *, HostType > &input, int n) const
Definition: load_balance.hpp:301
Kokkos::View< double **, Kokkos::LayoutRight, HostType > count_ptl_per_node_elec_main_ion(const Grid< DeviceType > &grid, const MagneticField< DeviceType > &magnetic_field, Plasma &plasma, bool kinetic_electrons)
Definition: count_ptl_per_node.cpp:29
double observed_load_imbalance
Definition: load_balance.hpp:33
void update_pol_decomp()
Definition: domain_decomposition.tpp:152
void update_model(const View< int *, CLayout, HostType > &current_partition)
Definition: load_balance.hpp:483
long long int get_total_n_ptl()
Definition: species.hpp:682
void f0_redistribute(Plasma &plasma, const DomainDecomposition< DeviceType > &pol_decomp, const Grid< DeviceType > &grid, const MagneticField< DeviceType > &magnetic_field, const VelocityGrid &vgrid, const View< int *, CLayout, HostType > &old_partition)
Definition: f0_redistribute.cpp:184
void set_new_partition(const Simulation< DeviceType > &sml, const Grid< DeviceType > &grid, const MagneticField< DeviceType > &magnetic_field, const VelocityGrid &vgrid, Plasma &plasma, DomainDecomposition< DeviceType > &pol_decomp, WeightingAlgorithm weighting_algorithm)
Definition: load_balance.hpp:525
#define TIMER(N, F)
Definition: timer_macro.hpp:24
void rebalance(const Simulation< DeviceType > &sml, const Grid< DeviceType > &grid, const MagneticField< DeviceType > &magnetic_field, const VelocityGrid &vgrid, Plasma &plasma, DomainDecomposition< DeviceType > &pol_decomp, ReweightOption reweight_option, WeightingAlgorithm weighting_algorithm=WeightingAlgorithm::Default)
Definition: load_balance.hpp:676
void use_namelist(const string &namelist)
Definition: NamelistReader.hpp:355
ConstraintOption
Definition: load_balance.hpp:279
void update_model(const View< int *, CLayout, HostType > &current_partition, const std::vector< double > &manual_times)
Definition: load_balance.hpp:490
double get_estimated_time_per_vertex(int i) const
Definition: load_balance.hpp:157
int my_period_rank
Definition: load_balance.hpp:42
int nnodes_on_plane
Number of nodes on local plane.
Definition: domain_decomposition.hpp:49
void calculate_load_imbalance(double f0_cost)
void initialize_model()
Definition: load_balance.hpp:198
WeightingAlgorithm default_weighting_algorithm
Definition: load_balance.hpp:290
View< int *, CLayout, HostType > proposed_partition
Which processors get which vertices.
Definition: load_balance.hpp:294
double get_time_since_previous_call()
Definition: load_balance.hpp:92
bool get_model_is_initialized() const
Definition: load_balance.hpp:177
UpdateMethod
Definition: load_balance.hpp:18
std::string region_name
Definition: load_balance.hpp:26
bool verbose
Definition: load_balance.hpp:28
bool f0_grid
Whether to use f0 grid.
Definition: sml.hpp:65
std::vector< LoadRegion > regions
Definition: load_balance.hpp:292
bool namelist_present(const string &namelist)
Definition: NamelistReader.hpp:351
UpdateMethod update_method
Definition: load_balance.hpp:46
bool model_is_initialized
Definition: load_balance.hpp:29
bool electron_on
Use kinetic electrons.
Definition: sml.hpp:64
void reset_cost_trackers()
void redistribute_load(const Simulation< DeviceType > &sml, const Grid< DeviceType > &grid, const MagneticField< DeviceType > &magnetic_field, const VelocityGrid &vgrid, Plasma &plasma, DomainDecomposition< DeviceType > &pol_decomp, const View< int *, CLayout, HostType > &old_partition)
Definition: load_balance.hpp:559
double get_largest_predicted_time(const View< int *, CLayout, HostType > &proposed_partition) const
Definition: load_balance.hpp:182
void initialize_model()
Definition: load_balance.hpp:496
T::value_type sum_view(const T &view)
Definition: view_arithmetic.hpp:56
bool recommend_proposed_partition()
Definition: load_balance.hpp:445
void exit_XGC(std::string msg)
Definition: globals.hpp:37
bool greedily_fill_partition(const View< double *, HostType > &weight, const View< double *, HostType > &constraint1, double target_weight_per_rank)
Definition: load_balance.hpp:311
int n_periods
How many repeating periods there are; in tokamaks this is planes.
Definition: load_balance.hpp:40
Definition: magnetic_field.F90:1
Definition: load_balance.hpp:16
void print_new_partition()
Definition: load_balance.hpp:475
WeightingAlgorithm
Definition: load_balance.hpp:272
int assess_whether_to_rebalance_load()
LoadBalance(NLReader::NamelistReader &nlr, const DomainDecomposition< DeviceType > &pol_decomp)
Definition: load_balance.hpp:613
View< double *, HostType > get_estimated_time_per_vertex() const
Definition: load_balance.hpp:161
Definition: plasma.hpp:14
double constraint1_max
Definition: load_balance.hpp:299
void set_weights(int *gvid0_pid, double *ptl_count, double *f0_node_cost)
View< double *, CLayout, HostType > count_all_ptl_per_node(const Grid< DeviceType > &grid, const MagneticField< DeviceType > &magnetic_field, Plasma &plasma)
Definition: count_ptl_per_node.cpp:71
ReweightOption
Definition: load_balance.hpp:283
View< double *, CLayout, HostType > f0_node_cost
Definition: plasma.hpp:38
void shift_all_species(Plasma &plasma, const Grid< DeviceType > &grid, const MagneticField< DeviceType > &magnetic_field, const DomainDecomposition< DeviceType > &pol_decomp, Shift::ShiftPh0 shift_ph0)
Definition: shift.cpp:301
void reset_timer()
Definition: load_balance.hpp:152
Definition: species.hpp:75
void update_model(const View< int *, CLayout, HostType > &current_partition, double manual_time=-1.0)
Definition: load_balance.hpp:205
bool model_is_initialized()
Definition: load_balance.hpp:502
std::vector< std::string > timer_names
Definition: load_balance.hpp:27
Definition: shift.hpp:10
bool pol_decomp
Use poloidal decomposition.
Definition: domain_decomposition.hpp:42
double get_observed_max_region_time() const
Definition: load_balance.hpp:169
Definition: load_balance.hpp:270
void touch_timers()
Definition: load_balance.hpp:116
Kokkos::ViewAllocateWithoutInitializing NoInit
Definition: space_settings.hpp:68
void update_model_no_history(const View< int *, CLayout, HostType > &current_partition, const View< double *, HostType > &all_periods_timings)
Definition: load_balance.hpp:48
double get_prediction_undershoot() const
Definition: load_balance.hpp:165
double get_observed_load_imbalance() const
Definition: load_balance.hpp:173
int get_max_n_ptl()
Definition: species.hpp:693
static int GPTLstop(const char *name)
Definition: timer_macro.hpp:10
Kokkos::View< int *, Kokkos::LayoutRight, HostType > gvid0_pid_h
Which processors get which vertices (host)
Definition: domain_decomposition.hpp:59