本文介绍了升压MPI挂起的处理方法，对大家解决问题具有一定的参考价值，需要的朋友们下面随着小编来一起学习吧！
问题描述
我有如下一个简单的异步消息传递程序。我在两台PC上用16个进程运行。这些过程都连接在一个torrus为4×4矩阵。因此，在主函数的开始，你会看到谁是一个过程的邻居。我所要做的是实现一个检查点算法。所以，每一个进程发送消息给所有的邻居当i = 5或10（假定我是时候）。当一个进程收到消息时，它启动为相同的处理的新irecv。但是，如果我是10不启动irecv，因为没有消息会后收到。在节目的最后，所有的进程等待未被接收的消息。
I have a simple asynchronous messaging program as below. I run it on two PCs with 16 processes. These processes are connected over a torrus as 4x4 matrix. So, at the beginning of main function you will see who are neighbours of a process. What I am trying to do is implementing a checkpointing algorithm. So every process sends a message to its all neighbors when i = 5 or 10 (assume that i is time). And when a process receives a message, it starts a new irecv for the same process. But if i is 10 it does not start an irecv, because no messages will be received after that. At the end of program, all processes wait for unreceived messages.
/* Demo_01_Main.cpp */
#include <boost/mpi.hpp>
#include <boost/serialization/string.hpp>
#include <string>
#include <iostream>
#include <fstream>
#include <map>
using namespace std;
class Packet{
friend class boost::serialization::access;
private:
int receiver;
int sender;
int data;
public:
Packet(){
receiver = 0;
sender = 0;
data = 0;
}
Packet(int receiver, int sender, int data){
this->receiver = receiver;
this->sender = sender;
this->data = data;
}
~Packet(){}
int getData() {
return data;
}
void setData(int data) {
this->data = data;
}
int getReceiver() {
return receiver;
}
void setReceiver(int receiver) {
this->receiver = receiver;
}
int getSender() {
return sender;
}
void setSender(int sender) {
this->sender = sender;
}
template<class Archive>
void serialize(Archive& ar, const unsigned int version) {
ar & receiver;
ar & sender;
ar & data;
}
string toString(){
stringstream ss;
ss << "Packet = [Data: " << data << ", Receiver: " << receiver << ", Sender: " << sender << "]";
return ss.str();
}
};
int rank;
void log(string str){
ofstream outfile;
stringstream logFileName;
logFileName << "log_" << rank << ".txt";
outfile.open(logFileName.str().c_str(), std::ios_base::app);
outfile << str;
outfile.close();
}
int main(int argc, char* argv[]){
map<int, boost::mpi::request> mpiReceiveRequest;
map<int, boost::mpi::request> mpiSendRequest;
map<int, Packet *> receivedData;
vector<int> neighbors;
boost::mpi::environment env(argc, argv);
boost::mpi::communicator world;
rank = world.rank();
if(rank == 0){
neighbors.push_back(1);
neighbors.push_back(3);
neighbors.push_back(4);
neighbors.push_back(5);
neighbors.push_back(7);
neighbors.push_back(12);
neighbors.push_back(13);
neighbors.push_back(15);
}
else if(rank == 1){
neighbors.push_back(0);
neighbors.push_back(2);
neighbors.push_back(4);
neighbors.push_back(5);
neighbors.push_back(6);
neighbors.push_back(12);
neighbors.push_back(13);
neighbors.push_back(14);
}
else if(rank == 2){
neighbors.push_back(1);
neighbors.push_back(3);
neighbors.push_back(5);
neighbors.push_back(6);
neighbors.push_back(7);
neighbors.push_back(13);
neighbors.push_back(14);
neighbors.push_back(15);
}
else if(rank == 3){
neighbors.push_back(0);
neighbors.push_back(2);
neighbors.push_back(4);
neighbors.push_back(6);
neighbors.push_back(7);
neighbors.push_back(12);
neighbors.push_back(14);
neighbors.push_back(15);
}
else if(rank == 4){
neighbors.push_back(0);
neighbors.push_back(1);
neighbors.push_back(3);
neighbors.push_back(5);
neighbors.push_back(7);
neighbors.push_back(8);
neighbors.push_back(9);
neighbors.push_back(11);
}
else if(rank == 5){
neighbors.push_back(0);
neighbors.push_back(1);
neighbors.push_back(2);
neighbors.push_back(4);
neighbors.push_back(6);
neighbors.push_back(8);
neighbors.push_back(9);
neighbors.push_back(10);
}
else if(rank == 6){
neighbors.push_back(1);
neighbors.push_back(2);
neighbors.push_back(3);
neighbors.push_back(5);
neighbors.push_back(7);
neighbors.push_back(9);
neighbors.push_back(10);
neighbors.push_back(11);
}
else if(rank == 7){
neighbors.push_back(0);
neighbors.push_back(2);
neighbors.push_back(3);
neighbors.push_back(4);
neighbors.push_back(6);
neighbors.push_back(8);
neighbors.push_back(10);
neighbors.push_back(11);
}
else if(rank == 8){
neighbors.push_back(4);
neighbors.push_back(5);
neighbors.push_back(7);
neighbors.push_back(9);
neighbors.push_back(11);
neighbors.push_back(12);
neighbors.push_back(13);
neighbors.push_back(15);
}
else if(rank == 9){
neighbors.push_back(4);
neighbors.push_back(5);
neighbors.push_back(6);
neighbors.push_back(8);
neighbors.push_back(10);
neighbors.push_back(12);
neighbors.push_back(13);
neighbors.push_back(14);
}
else if(rank == 10){
neighbors.push_back(5);
neighbors.push_back(6);
neighbors.push_back(7);
neighbors.push_back(9);
neighbors.push_back(11);
neighbors.push_back(13);
neighbors.push_back(14);
neighbors.push_back(15);
}
else if(rank == 11){
neighbors.push_back(4);
neighbors.push_back(6);
neighbors.push_back(7);
neighbors.push_back(8);
neighbors.push_back(10);
neighbors.push_back(12);
neighbors.push_back(14);
neighbors.push_back(15);
}
else if(rank == 12){
neighbors.push_back(0);
neighbors.push_back(1);
neighbors.push_back(3);
neighbors.push_back(8);
neighbors.push_back(9);
neighbors.push_back(11);
neighbors.push_back(13);
neighbors.push_back(15);
}
else if(rank == 13){
neighbors.push_back(0);
neighbors.push_back(1);
neighbors.push_back(2);
neighbors.push_back(8);
neighbors.push_back(9);
neighbors.push_back(10);
neighbors.push_back(12);
neighbors.push_back(14);
}
else if(rank == 14){
neighbors.push_back(1);
neighbors.push_back(2);
neighbors.push_back(3);
neighbors.push_back(9);
neighbors.push_back(10);
neighbors.push_back(11);
neighbors.push_back(13);
neighbors.push_back(15);
}
else if(rank == 15){
neighbors.push_back(0);
neighbors.push_back(2);
neighbors.push_back(3);
neighbors.push_back(8);
neighbors.push_back(10);
neighbors.push_back(11);
neighbors.push_back(12);
neighbors.push_back(14);
}
for(int i=0; i<8; i++){
Packet * packet = new Packet();
receivedData[neighbors[i]] = packet;
mpiReceiveRequest[neighbors[i]] = world.irecv(neighbors[i], 100, *packet);
}
for(int i=1; i<=10; i++){
if(i%5 == 0){ // Checkpoint time
for(int j=0; j<8; j++){
Packet * p = new Packet(neighbors[j], rank, i);
mpiSendRequest[neighbors[j]] = world.isend(neighbors[j], 100, *p);
log("Sending: ");
log(p->toString());
log("\n");
}
}
for(int j=0; j<8; j++){
if(mpiReceiveRequest[neighbors[j]].test()){
Packet * p = receivedData[neighbors[j]];
log("Received: ");
log(receivedData[neighbors[j]]->toString());
log("\n");
if(p->getData() != 10){
Packet * packet = new Packet();
receivedData[neighbors[j]] = packet;
mpiReceiveRequest[neighbors[j]] = world.irecv(neighbors[j], 100, *packet);
}
}
}
}
for(int i=0; i<8; i++){
stringstream ss;
ss << " Wait from: " << neighbors[i] << endl;
log(ss.str());
mpiReceiveRequest[neighbors[i]].wait();
log("Received: ");
log(receivedData[neighbors[i]]->toString());
log("\n");
}
stringstream ss;
ss << rank << " is done" << endl;
log(ss.str());
return 0;
}
的问题是，它在等待命令挂起。同时，一些收到的消息中包含的数据废话。例如，处理7的输出文件是象下面这样：
The problem is that it hangs on at the wait command. As well, some of the received messages contains nonsense data. For instance, the output file of process 7 is like below:
Received: Packet = [Data: 5, Receiver: 7, Sender: 10]
Received: Packet = [Data: 5, Receiver: 7, Sender: 11]
Received: Packet = [Data: 5, Receiver: 7, Sender: 0]
Received: Packet = [Data: 5, Receiver: 7, Sender: 4]
Sending: Packet = [Data: 5, Receiver: 0, Sender: 7]
Sending: Packet = [Data: 5, Receiver: 2, Sender: 7]
Sending: Packet = [Data: 5, Receiver: 3, Sender: 7]
Sending: Packet = [Data: 5, Receiver: 4, Sender: 7]
Sending: Packet = [Data: 5, Receiver: 6, Sender: 7]
Sending: Packet = [Data: 5, Receiver: 8, Sender: 7]
Sending: Packet = [Data: 5, Receiver: 10, Sender: 7]
Sending: Packet = [Data: 5, Receiver: 11, Sender: 7]
Wait from: 0
Received: Packet = [Data: 537985024, Receiver: 0, Sender: 0]
Wait from: 2
我想不出问题出在哪里。
I could not figure out where the problem is.
推荐答案
问题是在调用wait命令是前测试，并顺利通过了要求。其实，我意识到，这也是我的previous问题之一的原因：Waiting请求已已经完成
The problem was on calling wait command for the request which is tested before and passed. Actually, I realized that it is also the reason of one of my previous questions: Waiting requests which have been already completed
所以，我修改了code如下，现在它运行：
So, I modified my code as below, and now it runs:
/* ...
else if(rank == 15){
neighbors.push_back(0);
neighbors.push_back(2);
neighbors.push_back(3);
neighbors.push_back(8);
neighbors.push_back(10);
neighbors.push_back(11);
neighbors.push_back(12);
neighbors.push_back(14);
}
...After assigning neighbors
*/
for(int i=0; i<8; i++){
Packet * packet = new Packet();
receivedData[neighbors[i]] = packet;
mpiReceiveRequest[i] = world.irecv(neighbors[i], 100, *packet);
}
vector<int> completed;
for(int i=1; i<=10; i++){
if(i%5 == 0){ // Checkpoint time
for(int j=0; j<8; j++){
Packet * p = new Packet(neighbors[j], rank, i);
mpiSendRequest.push_back(world.isend(neighbors[j], 100, *p));
log("Sending: ");
log(p->toString());
log("\n");
}
}
for(int j=0; j<8; j++){
vector<int>::iterator it = completed.begin();
bool passed = false;
while(it != completed.end()){
if(*it == j){
passed = true;
break;
}
it++;
}
if(!passed){
if(mpiReceiveRequest[j].test()){
completed.push_back(j);
Packet * p = receivedData[neighbors[j]];
if(p->getData() != 10){
Packet * packet = new Packet();
receivedData[neighbors[j]] = packet;
mpiReceiveRequest[j] = world.irecv(neighbors[j], 100, *packet);
completed.pop_back();
}
}
}
}
}
vector<boost::mpi::request> reqs;
for(int i=0; i<8; i++){
vector<int>::iterator it = completed.begin();
bool passed = false;
while(it != completed.end()){
if(*it == i){
passed = true;
break;
}
it++;
}
if(!passed){
mpiReceiveRequest[i].wait();
reqs.push_back(mpiReceiveRequest[i]);
}
}
stringstream ss;
ss << rank << " is done" << endl;
log(ss.str());
return 0;
} // End of main
在结束
的主要区别是存储在一个载体完成的请求，而不是等着他们。
The main difference is storing the completed requests in a vector and not waiting them at the end.
这篇关于升压MPI挂起的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持！