void ThreadPool::submitBarrier() {
//Promises are not copyable, so we save a pointer and delete it later, after the barrier.
auto promise = new std::promise<void>();
std::shared_future<void> future(promise->get_future());
auto counter = new std::atomic<int>();
counter->store(0);
int goal = thread_count;
auto barrierJob = [counter, promise, future, goal] () {
int currentCounter = counter->fetch_add(1); //increment by one and get the current counter.
if(currentCounter == goal-1) { //we're finished.
promise->set_value();
delete promise; //we're done, this isn't needed anymore.
//We got here, so we're the last one to manipulate the counter.
delete counter;
}
else {
//Otherwise, we wait for all the other barriers.
future.wait();
}
};
for(int i = 0; i < thread_count; i++) submitJob(barrierJob);
}
bool CommonBfJlModules::dispatch(NodePtr pNode)
{
assert(m_pBfsHandler != nullptr);
assert(m_pGameHandler != nullptr);
std::ostringstream oss;
if (m_pBfsHandler->isRunningJob(pNode)) {
oss << "Error : the node to dispathch is running job" << std::endl;
std::cerr << oss.str() << "> " << std::flush;
pNode->appendComment(oss.str());
return false;
} else if (m_pBfsHandler->isFlagged(pNode)) {
oss << "Error : the node to dispatch is flagged" << std::endl;
std::cerr << oss.str() << "> " << std::flush;
pNode->appendComment(oss.str());
return false;
}
std::string sAppName = m_pGameHandler->getAppName();
std::string sAppVersion = m_pGameHandler->getAppVersion();
std::string sArgument = m_pGameHandler->prepareJobCommands(pNode);
int iJId = submitJob(pNode, sAppName, sAppVersion, sArgument);
oss.str("");
oss << "Submit job ID " << iJId << ": " << sArgument << std::endl;
pNode->appendComment(oss.str());
if (JobLevelConfigure::g_configure.bUseJobLog) {
std::ofstream fout(JobLevelConfigure::g_configure.sJobLogFile.c_str(), std::ios::app);
fout << oss.str();
fout.close();
}
m_nDoingJobs++;
m_nTotalJobs++;
return true;
}
/*
* This file represents the client half of the JCEP communication. It
* operates in much the same way as the RMI skeleton expect for two important
* points:
* 1) It doesn't use JRMP. Instead it uses JCEP. JCEP is a much simpler and
* more expressive protocol for this specific task. This means that all the
* complicated unwinding of serialized Java objects that the RMI skeleton had
* are no longer necesasary. This JCEP skeleton does not have to understand
* one lick of Java.
* 2) It is a two-way communication. The RMI skeleton just feed a job to the
* server and read the response, which either said it was received or not. It
* did not continue running for the duration of the job. That means that SGE
* couldn't track the job past its submission. That also meant that problems
* with the job that came after execution started had to be passed back to the
* ComputeProxy is be logged. Now the skeleton runs through the entire
* duration of the job. Any messages or errors generated as the job runs are
* sent to the skeleton, which will print them out to stdout or stderr so that
* SGE can log them. That means that all information from the running of the
* job is in the same SGE-generated files. It also means that the job continues
* to be trackable by SGE throughout its duration. This two-way communication
* also means that the JCEP skeleton has to be multi-threaded.
*
* Only one JCEPskeleton may be allowed to attach to a given job. The reason is
* that a failed command for one will terminate them all. See the comments on
* listenerThread() about line 600 for more info.
*/
int main (int argc, char** argv) {
int serverPort, threadStatus, code;
char serverIP[32];
pthread_t threadId, sigThreadId;
//Process arguments
processArgs (argc, argv);
//Set the buffering on STDOUT to line buffering
setlinebuf (stdout);
//Connect to server
if (debug) {
printf ("**DEBUG: Beginning skeleton...\n");
printf ("**DEBUG: Creating socket to JCEP server...\n");
}
// Open a socket to RMI registry
if (createConnection ("127.0.0.1", PORT) < 0) {
fprintf (stderr, "Unable to establish connection to JCEP server:\n\t%s\n", errorMessage);
exit (1);
}
if (debug) {
printf ("**DEBUG: Connection established.\n");
}
/* The order of the following events is purposeful. The signal thread must
* start before the listener thread so that the listener thread can inherit
* the signal mask set in startSignalThread(). The signal thread has to
* start after the command is sent because otherwise there could be
* synchronization problems with the main thread and the signal thread both
* trying to send a command to the server at the same time. Starting the
* signal thread after the command does leave a small hole, where if SGE
* tried to cancel or suspend the job immediately after starting it, the
* signal could come before the signal thread is ready. This should be fixed
* in the future with some explicit synchronization. */
if (taskType == SUBMIT_JOB) {
/* Send job to server */
submitJob ();
}
else if (taskType == SHUTDOWN) {
shutdownServer ();
}
else if (taskType == CHECKPOINT_JOB) {
checkpointJob (jobId);
}
else if (taskType == CANCEL_JOB) {
cancelJob (jobId);
}
else if (taskType == REGISTER) {
listenToJob (jobId);
}
else if (taskType == SUSPEND) {
suspendJob (jobId);
}
else if (taskType == RESUME) {
resumeJob (jobId);
}
/* Start up a thread to handle SIGUSR1 (pending suspend), SIGUSR2 (pending
* cancel), and SIGCONT (just resumed). */
if ((code = startSignalThread (&sigThreadId)) != 0) {
fprintf (stderr, "Unable to start signal thread: error code %d.\n", code);
exit (1);
}
/* Start listener thread */
if ((code = startListenerThread (&threadId)) != 0) {
fprintf (stderr, "Unable to start listener thread: error code %d.\n", code);
exit (1);
}
/* Wait for a job finished notice */
pthread_join (threadId, (void**)&threadStatus);
if (debug) {
printf ("**DEBUG: Listener thread has exited.\n");
}
if (threadStatus < 0) {
fprintf (stderr, "Listener thread received the following error:\n\t%s\n", errorMessage);
exit (1);
}
/* Exit upon receipt of job finished notice */
return (EXIT_SUCCESS);
}