Exchanger and MPI Compile Flag #50
Conversation
seed generation to the point before the seed is actually placed in the program to ensure that the different mpi processes are actually doing something different.
and out of the logic of other classes
Codecov Report
@@ Coverage Diff @@
## master #50 +/- ##
==========================================
- Coverage 87.1% 58.6% -28.51%
==========================================
Files 178 134 -44
Lines 16266 10621 -5645
Branches 7849 3300 -4549
==========================================
- Hits 14168 6224 -7944
- Misses 2097 4397 +2300
+ Partials 1 0 -1
Continue to review full report at Codecov.
|
| @@ -380,7 +381,7 @@ struct EvolutionaryParameters { | |||
| double gamma; | |||
| size_t edge_frequency_amount; | |||
| bool dynamic_population_size; | |||
| double dynamic_population_amount_of_time; | |||
| float dynamic_population_amount_of_time; | |||
| equal_sequential_time, | ||
| equal_to_the_number_of_mpi_processes, | ||
| as_usual | ||
| }; |
There was a problem hiding this comment.
We need better names for these 3 strategies.
There was a problem hiding this comment.
as_usual = 15_percent_of_total_time
equal_sequential_time = 15_percent_of_total_time_times_num_procs
equal_to_the_number_of_mpi_processes = equal_to_the_number_of_mpi_processes
- add some comments that briefly describe the strategies
kahypar/partition/evo_partitioner.h
Outdated
| DBG << " "; | ||
| } | ||
| DBG << ""; | ||
| inline std::string preface(const Context& context) { |
There was a problem hiding this comment.
This should probably be a method of the communicator.
| @@ -55,12 +63,12 @@ class Population { | |||
| } | |||
| } | |||
| inline size_t forceInsert(Individual&& individual, const size_t position) { | |||
| DBG << V(position) << V(individual.fitness()); | |||
| DBG << " MPIRank " << _mpi_rank << ":" << V(position) << V(individual.fitness()); | |||
There was a problem hiding this comment.
If you make preface a method of the communicator, it can also be used here.
| evo_part.partition(hypergraph, context); | ||
| std::vector<double> times = Timer::instance().evolutionaryResult().evolutionary; | ||
| double total_time = Timer::instance().evolutionaryResult().total_evolutionary; | ||
| ASSERT_GT(total_time, context.partition.time_limit); | ||
| ASSERT_LT(total_time - times.at(times.size() - 1), context.partition.time_limit); | ||
| // TODO verify |
| } | ||
| inline void broadcastPopulationSize(Context& context) { | ||
| MPI_Bcast(&context.evolutionary.population_size, 1, MPI_INT, 0, MPI_COMM_WORLD); | ||
| MPI_Barrier(MPI_COMM_WORLD); |
There was a problem hiding this comment.
Why do you need this barrier?
| } | ||
|
|
||
|
|
||
| inline int getRandomOpenTarget() { |
There was a problem hiding this comment.
This methods seems to be overkill. Why don't you just choose a random number and check whether or not _individual_already_sent_to[current_target] is false?
|
|
||
|
|
||
| // Permutates the vector so that no element is left on the original position in the vector. | ||
| inline void degenerate(std::vector<int>& vector) { |
There was a problem hiding this comment.
This method needs a better name. Also, why didn't you just use: https://en.cppreference.com/w/cpp/algorithm/random_shuffle
| int sending_to = permutation_of_mpi_process_numbers[_rank]; | ||
| int receiving_from = findCorrespondingSender(permutation_of_mpi_process_numbers); | ||
|
|
||
| std::vector<PartitionID> outgoing_partition = population.individualAt(population.randomIndividual()).partition(); |
There was a problem hiding this comment.
Use std::vector& to prevent unnecessary vector copy - or do you really need a copy here because of async send?
| &received_partition_vector[0], 1, _MPI_Partition, receiving_from, 0, _communicator, &st); | ||
|
|
||
|
|
||
| hg.reset(); |
There was a problem hiding this comment.
again, reset is not necessary because it is called in setPartition
| std::vector<PartitionID> outgoing_partition = population.individualAt(population.randomIndividual()).partition(); | ||
| std::vector<PartitionID> received_partition_vector; | ||
| received_partition_vector.resize(hg.initialNumNodes()); | ||
| DBG << preface() << "receiving from " << receiving_from << " sending to " << sending_to << "quick_start"; |
There was a problem hiding this comment.
allocate vector directly with initialNumNode entries
… in command line options
…omize, introduced a local partition buffer
and out of the logic of other classes
… in command line options
…omize, introduced a local partition buffer
Introduces the MPI exchange protocol and the necessary stuff to encapsulate MPI