如何解决关于MPI_Isend
我正在使用MPI库在master和worker之间传递对象数组。由于对象众多,因此我在循环中使用MPI_Isend
,然后使用MPI_Recv
。除了最后一个MPI_Recv
仅在MPI_Recv
接收到最后发送的元素MPI_Isent
之外,其他所有内容都正常运行,此外,该代码将无任何错误地终止。
发送的输入是:
1 2 3 0.4 0 4 5 1
2 4 7 0.678 0 8 1
....
5 8 7 0.56 0 6 1
,收到的输入是:
5 8 7 0.56 0 6 1
5 8 7 0.56 0 6 1
5 8 7 0.56 0 6 1
.... (always the same)
我的问题很简单:为什么?我感觉在MPI传输中有些东西被覆盖了,但是我不知道在哪里以及为什么。我的代码的第一部分似乎还可以,我的问题在最后一部分。
编辑
根据评论,问题在于没有MPI_Waitall()
验证完成。我必须承认无法理解在何处插入任何MPI_Wait
或MPI_Waitall
。
main.cpp
MPI_Datatype mpi_gene = MPI_Send_Genes();
MPI_Request reqs_6[size_req],reqs_7[size_req];
MPI_Status status_7[size_req];
if(rank==0){
for(int r=1; r < com; r++){
MPI_Isend(&num_orgs,1,MPI_INT,r,parallel_0,&reqs_0[r]);
for(int org = 0; org < NEAT::pop_size; org++){
//send some genome information
genome_info g_info;
g_info.generation = pop->organisms[org]->generation;
MPI_Isend(&g_info,mpi_genome_info,org,parallel_1,&reqs_1[org]);
//define the genome
NEAT::Genome* sent_genome = pop->organisms[org]->gnome;
//third the genes
std::vector<NEAT::Gene*> gen_genes = sent_genome->genes;
int num_genes = gen_genes.size();
struct genes sent_genes[num_genes];
for(int id = 0; id < num_genes; id++){
if(gen_genes[id]->lnk->linktrait==0){
sent_genes[id].trait_id = 0;
}else{
sent_genes[id].trait_id = gen_genes[id]->lnk->linktrait->trait_id;
}
sent_genes[id].in_node_id = gen_genes[id]->lnk->in_node->node_id;
sent_genes[id].out_node_id = gen_genes[id]->lnk->out_node->node_id;
sent_genes[id].weight = gen_genes[id]->lnk->weight;
sent_genes[id].recurrent = gen_genes[id]->lnk->is_recurrent;
sent_genes[id].innovation_num = gen_genes[id]->innovation_num;
sent_genes[id].mutation_num = gen_genes[id]->mutation_num;
sent_genes[id].enable = gen_genes[id]->enable;
if(id==3){
std::cout <<
sent_genes[id].in_node_id << " " <<
sent_genes[id].out_node_id <<" " <<
sent_genes[id].weight <<" " <<
sent_genes[id].recurrent <<" " <<
sent_genes[id].innovation_num <<" " <<
sent_genes[id].mutation_num <<" " <<
sent_genes[id].enable <<" " <<
std::endl;
}
}
MPI_Isend(&num_genes,parallel_6,&reqs_6[org]);
MPI_Isend(&sent_genes,num_genes,mpi_gene,parallel_7,&reqs_7[org]);
}
}
}
std::cout << "--------------" << std::endl;
if(rank!=0){
std::vector<NEAT::Organism*> local_pop;
int generation;
MPI_Recv(&num_orgs,&status);
for(int org = 0; org < num_orgs ; org++){
//receive genome information
genome_info rcvd_genome_info;
MPI_Recv(&rcvd_genome_info,&status);
generation = rcvd_genome_info.generation;
//receive genes
int num_rcvd_genes;
MPI_Recv(&num_rcvd_genes,&status);
genes rcvd_genes[num_rcvd_genes];
MPI_Recv(&rcvd_genes,num_rcvd_genes,&status);
MPI_Waitall(size_req,&reqs_7[size_req],MPI_STATUSES_IGNORE);
std::cout << num_rcvd_genes << std::endl; //this is ok
std::vector<NEAT::Gene*> gen_genes;
for(int id = 0; id < num_rcvd_genes; id++){
genes p_gene = rcvd_genes[id];
if(id==3){//PROBLEM HERE
std::cout << id << " <- "<<
p_gene.in_node_id << " " <<
p_gene.out_node_id <<" " <<
p_gene.weight <<" " <<
p_gene.recurrent <<" " <<
p_gene.innovation_num <<" " <<
p_gene.mutation_num <<" " <<
p_gene.enable <<" " <<
std::endl;
}
}
}
MPI_Waitall(size_req,MPI_STATUSES_IGNORE);
}
send_genes.cpp
MPI_Datatype MPI_Send_Genes(){
int nitems = 8;
int gene_blocklengths[nitems] = {1,1};
MPI_Datatype gene_types[nitems] = {MPI_INT,MPI_DOUBLE,MPI_C_BOOL,MPI_C_BOOL};
MPI_Datatype mpi_gene;
MPI_Aint gene_offsets[nitems];
gene_offsets[0] = offsetof(genes,trait_id);
gene_offsets[1] = offsetof(genes,in_node_id);
gene_offsets[2] = offsetof(genes,out_node_id);
gene_offsets[3] = offsetof(genes,weight);
gene_offsets[4] = offsetof(genes,recurrent);
gene_offsets[5] = offsetof(genes,innovation_num);
gene_offsets[6] = offsetof(genes,mutation_num);
gene_offsets[7] = offsetof(genes,enable);
MPI_Type_create_struct(nitems,gene_blocklengths,gene_offsets,gene_types,&mpi_gene);
MPI_Type_commit(&mpi_gene);
return mpi_gene;
}
send_genes.h
struct genes{
int trait_id;
int in_node_id;
int out_node_id;
double weight;
bool recurrent;
double innovation_num;
double mutation_num;
bool enable;
};
解决方法
由于Hristo Iliev和Gilles Gouaillardet的提示,我注意到我并不能确保每个请求都已正确发起和发送。如果我的理解是正确的,我只需要在每个memcpy
之后添加一个MPI_Wait
,就可以正常使用了。但是,我必须承认仍然不理解为什么仅复制和发送最后一个值。
MPI_Isend
版权声明:本文内容由互联网用户自发贡献,该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 dio@foxmail.com 举报,一经查实,本站将立刻删除。