int getAprunArg(aprun_arg_t argt) {
switch (argt) {
case aprun_cc:
return -1; // string arg/user provides this
case aprun_n:
return -1; // user provides this
case aprun_d:
return getCoresPerLocale();
case aprun_N:
return getLocalesPerNode();
case aprun_j:
return getCPUsPerCU();
default:
return -1;
}
}
// create the command that will actually launch the program and
// create any files needed for the launch like the batch script
static char* chpl_launch_create_command(int argc, char* argv[],
int32_t numLocales) {
int i;
int size;
char baseCommand[256];
char* command;
FILE* slurmFile, *expectFile;
char* account = getenv("CHPL_LAUNCHER_ACCOUNT");
char* constraint = getenv("CHPL_LAUNCHER_CONSTRAINT");
char* outputfn = getenv("CHPL_LAUNCHER_SLURM_OUTPUT_FILENAME");
char* basenamePtr = strrchr(argv[0], '/');
pid_t mypid;
// command line walltime takes precedence over env var
if (!walltime) {
walltime = getenv("CHPL_LAUNCHER_WALLTIME");
}
// command line nodelist takes precedence over env var
if (!nodelist) {
nodelist = getenv("CHPL_LAUNCHER_NODELIST");
}
if (basenamePtr == NULL) {
basenamePtr = argv[0];
} else {
basenamePtr++;
}
chpl_compute_real_binary_name(argv[0]);
if (debug) {
mypid = 0;
} else {
mypid = getpid();
}
// set the filenames
sprintf(expectFilename, "%s%d", baseExpectFilename, (int)mypid);
sprintf(slurmFilename, "%s%d", baseSBATCHFilename, (int)mypid);
// if were running a batch job
if (getenv("CHPL_LAUNCHER_USE_SBATCH") != NULL || generate_sbatch_script) {
// open the batch file and create the header
slurmFile = fopen(slurmFilename, "w");
fprintf(slurmFile, "#!/bin/sh\n\n");
// set the job name
fprintf(slurmFile, "#SBATCH --job-name=Chpl-%.10s\n", basenamePtr);
// suppress informational messages, will still display errors
fprintf(slurmFile, "#SBATCH --quiet\n");
// request the number of locales, with 1 task per node, and number of cores
// cpus-per-task. We probably don't need --nodes and --ntasks specified
// since 1 task-per-node with n --tasks implies -n nodes
fprintf(slurmFile, "#SBATCH --nodes=%d\n", numLocales);
fprintf(slurmFile, "#SBATCH --ntasks=%d\n", numLocales);
fprintf(slurmFile, "#SBATCH --ntasks-per-node=%d\n", procsPerNode);
fprintf(slurmFile, "#SBATCH --cpus-per-task=%d\n", getCoresPerLocale());
//request exclusive access to nodes
fprintf(slurmFile, "#SBATCH --exclusive\n");
// Set the walltime if it was specified
if (walltime) {
fprintf(slurmFile, "#SBATCH --time=%s\n", walltime);
}
// Set the nodelist if it was specified
if (nodelist) {
fprintf(slurmFile, "#SBATCH --nodelist=%s\n", nodelist);
}
// If needed a constraint can be specified with the env var CHPL_LAUNCHER_CONSTRAINT
if (constraint) {
fprintf(slurmFile, "#SBATCH --constraint=%s\n", constraint);
}
// set the account name if one was provided
if (account && strlen(account) > 0) {
fprintf(slurmFile, "#SBATCH --account=%s\n", account);
}
// set the output name to either the user specified
// or to the binaryName.<jobID>.out if none specified
if (outputfn!=NULL) {
fprintf(slurmFile, "#SBATCH --output=%s\n", outputfn);
}
else {
fprintf(slurmFile, "#SBATCH --output=%s.%%j.out\n", argv[0]);
}
// add the srun command
fprintf(slurmFile, "srun %s ", chpl_get_real_binary_name());
// add any arguments passed to the launcher to the binary
for (i=1; i<argc; i++) {
fprintf(slurmFile, " '%s'", argv[i]);
}
fprintf(slurmFile, "\n");
// close the batch file and change permissions
fclose(slurmFile);
chmod(slurmFilename, 0755);
if (generate_sbatch_script) {
fprintf(stdout, "SBATCH script written to '%s'\n", slurmFilename);
}
// the baseCommand is what will call the batch file
// that was just created
sprintf(baseCommand, "sbatch %s\n", slurmFilename);
}
// else we're running an interactive job
else {
// expect is used to launch an interactive job
// create the file and set some things for expect
expectFile = fopen(expectFilename, "w");
fprintf(expectFile, "set timeout -1\n");
fprintf(expectFile, "set prompt \"(%%|#|\\\\$|>) $\"\n");
// create a silent salloc command
fprintf(expectFile, "spawn -noecho srun ");
// set the job name
fprintf(expectFile, "--job-name=CHPL-%.10s ",basenamePtr);
// suppress informational messages, will still display errors
fprintf(expectFile, "--quiet ");
// request the number of locales, with 1 task per node, and number of cores
// cpus-per-task. We probably don't need --nodes and --ntasks specified
// since 1 task-per-node with n --tasks implies -n nodes
fprintf(expectFile, "--nodes=%d ",numLocales);
fprintf(expectFile, "--ntasks=%d ", numLocales);
fprintf(expectFile, "--ntasks-per-node=%d ", procsPerNode);
fprintf(expectFile, "--cpus-per-task=%d ", getCoresPerLocale());
// request exclusive access
fprintf(expectFile, "--exclusive ");
// Set the walltime if i was specified
if (walltime) {
fprintf(expectFile, "--time=%s ", walltime);
}
// Set the walltime if it was specified
if (nodelist) {
fprintf(expectFile, "--nodelist=%s ", nodelist);
}
// set any constraints
if (constraint) {
fprintf(expectFile, " --constraint=%s ", constraint);
}
// set the account name if one was provided
if (account && strlen(account) > 0) {
fprintf(expectFile, "--account=%s ", account);
}
// the actual srun command
fprintf(expectFile, "%s", chpl_get_real_binary_name());
// add any arguments passed to the launcher to the binary
for (i=1; i<argc; i++) {
fprintf(expectFile, " %s", argv[i]);
}
fprintf(expectFile, "\n\n");
// do some things required for expect and close the file
fprintf(expectFile, "interact -o -re $prompt {return}\n");
fclose(expectFile);
// the baseCommand is what will call the expect file
sprintf(baseCommand, "expect %s", expectFilename);
}
// copy baseCommand into command and return it
size = strlen(baseCommand) + 1;
command = chpl_mem_allocMany(size, sizeof(char), CHPL_RT_MD_COMMAND_BUFFER, -1, "");
sprintf(command, "%s", baseCommand);
if (strlen(command)+1 > size) {
chpl_internal_error("buffer overflow");
}
return command;
}
// create the command that will actually launch the program and
// create any files needed for the launch like the batch script
static char* chpl_launch_create_command(int argc, char* argv[],
int32_t numLocales) {
int i;
int size;
char baseCommand[MAX_COM_LEN];
char* command;
FILE* slurmFile;
char* account = getenv("CHPL_LAUNCHER_ACCOUNT");
char* constraint = getenv("CHPL_LAUNCHER_CONSTRAINT");
char* outputfn = getenv("CHPL_LAUNCHER_SLURM_OUTPUT_FILENAME");
char* basenamePtr = strrchr(argv[0], '/');
pid_t mypid;
// For programs with large amounts of output, a lot of time can be
// spent syncing the stdout buffer to the output file. This can cause
// tests to run extremely slow and can cause stdout and stderr to
// become mixed in odd ways since stdout is buffered but stderr isn't.
// To alleviate this problem (and to allow accurate external timings
// of tests) this allows the output to be "buffered" to <tmpDir> and
// copied once the job is done.
//
// Note that this should work even for multi-locale tests since all
// the output is piped through a single node.
//
// The *NoFmt versions are the same as the regular version, except
// that instead of using slurms output formatters, they use the
// corresponding env var. e.g. you have to use '--output=%j.out to
// have the output file be <jobid>.out, but when we copy the tmp file
// to the real output file, the %j and other formatters aren't
// available so we have to use the equivalent slurm env var
// (SLURM_JOB_ID.) The env vars can't be used when specifying --output
// because they haven't been initialized yet
char* bufferStdout = getenv("CHPL_LAUNCHER_SLURM_BUFFER_STDOUT");
const char* tmpDir = getTmpDir();
char stdoutFile [MAX_COM_LEN];
char stdoutFileNoFmt [MAX_COM_LEN];
char tmpStdoutFile [MAX_COM_LEN];
char tmpStdoutFileNoFmt [MAX_COM_LEN];
// command line walltime takes precedence over env var
if (!walltime) {
walltime = getenv("CHPL_LAUNCHER_WALLTIME");
}
// command line nodelist takes precedence over env var
if (!nodelist) {
nodelist = getenv("CHPL_LAUNCHER_NODELIST");
}
// command line partition takes precedence over env var
if (!partition) {
partition = getenv("CHPL_LAUNCHER_PARTITION");
}
// command line exclude takes precedence over env var
if (!exclude) {
exclude = getenv("CHPL_LAUNCHER_EXCLUDE");
}
if (basenamePtr == NULL) {
basenamePtr = argv[0];
} else {
basenamePtr++;
}
chpl_compute_real_binary_name(argv[0]);
if (debug) {
mypid = 0;
} else {
mypid = getpid();
}
// Elliot, 12/02/14: TODO we have a bunch of similar commands to build up the
// interactive and batch versions. It would be nicer to build up the commands
// and postprocess depending on interactive vs batch. As in build up "--quiet
// --nodes ..." and afterwards split on ' ' and then add #SBATCH and a
// newline for batch mode and leave it as is for interactive"
// if were running a batch job
if (getenv("CHPL_LAUNCHER_USE_SBATCH") != NULL || generate_sbatch_script) {
// set the sbatch filename
sprintf(slurmFilename, "%s%d", baseSBATCHFilename, (int)mypid);
// open the batch file and create the header
slurmFile = fopen(slurmFilename, "w");
fprintf(slurmFile, "#!/bin/sh\n\n");
// set the job name
fprintf(slurmFile, "#SBATCH --job-name=Chpl-%.10s\n", basenamePtr);
// suppress informational messages, will still display errors
fprintf(slurmFile, "#SBATCH --quiet\n");
// request the number of locales, with 1 task per node, and number of cores
// cpus-per-task. We probably don't need --nodes and --ntasks specified
// since 1 task-per-node with n --tasks implies -n nodes
fprintf(slurmFile, "#SBATCH --nodes=%d\n", numLocales);
fprintf(slurmFile, "#SBATCH --ntasks=%d\n", numLocales);
fprintf(slurmFile, "#SBATCH --ntasks-per-node=%d\n", procsPerNode);
fprintf(slurmFile, "#SBATCH --cpus-per-task=%d\n", getCoresPerLocale());
//request exclusive access to nodes
fprintf(slurmFile, "#SBATCH --exclusive\n");
// Set the walltime if it was specified
if (walltime) {
fprintf(slurmFile, "#SBATCH --time=%s\n", walltime);
}
// Set the nodelist if it was specified
if (nodelist) {
fprintf(slurmFile, "#SBATCH --nodelist=%s\n", nodelist);
}
// Set the partition if it was specified
if (partition) {
fprintf(slurmFile, "#SBATCH --partition=%s\n", partition);
}
// Set the exclude list if it was specified
if (exclude) {
fprintf(slurmFile, "#SBATCH --exclude=%s\n", exclude);
}
// If needed a constraint can be specified with the env var CHPL_LAUNCHER_CONSTRAINT
if (constraint) {
fprintf(slurmFile, "#SBATCH --constraint=%s\n", constraint);
}
// set the account name if one was provided
if (account && strlen(account) > 0) {
fprintf(slurmFile, "#SBATCH --account=%s\n", account);
}
// set the output file name to either the user specified
// name or to the binaryName.<jobID>.out if none specified
if (outputfn != NULL) {
sprintf(stdoutFile, "%s", outputfn);
sprintf(stdoutFileNoFmt, "%s", outputfn);
}
else {
sprintf(stdoutFile, "%s.%s.out", argv[0], "%j");
sprintf(stdoutFileNoFmt, "%s.%s.out", argv[0], "$SLURM_JOB_ID");
}
// We have slurm use the real output file to capture slurm errors/timeouts
// We only redirect the program output to the tmp file
fprintf(slurmFile, "#SBATCH --output=%s\n", stdoutFile);
// If we're buffering the output, set the temp output file name.
// It's always <tmpDir>/binaryName.<jobID>.out.
if (bufferStdout != NULL) {
sprintf(tmpStdoutFile, "%s/%s.%s.out", tmpDir, argv[0], "%j");
sprintf(tmpStdoutFileNoFmt, "%s/%s.%s.out", tmpDir, argv[0], "$SLURM_JOB_ID");
}
// add the srun command and the (possibly wrapped) binary name.
fprintf(slurmFile, "srun --kill-on-bad-exit %s %s ",
chpl_get_real_binary_wrapper(), chpl_get_real_binary_name());
// add any arguments passed to the launcher to the binary
for (i=1; i<argc; i++) {
fprintf(slurmFile, "'%s' ", argv[i]);
}
// buffer program output to the tmp stdout file
if (bufferStdout != NULL) {
fprintf(slurmFile, "&> %s", tmpStdoutFileNoFmt);
}
fprintf(slurmFile, "\n");
// After the job is run, if we buffered stdout to <tmpDir>, we need
// to copy the output to the actual output file. The <tmpDir> output
// will only exist on one node, ignore failures on the other nodes
if (bufferStdout != NULL) {
fprintf(slurmFile, "cat %s >> %s\n", tmpStdoutFileNoFmt, stdoutFileNoFmt);
fprintf(slurmFile, "rm %s &> /dev/null\n", tmpStdoutFileNoFmt);
}
// close the batch file and change permissions
fclose(slurmFile);
chmod(slurmFilename, 0755);
if (generate_sbatch_script) {
fprintf(stdout, "SBATCH script written to '%s'\n", slurmFilename);
}
// the baseCommand is what will call the batch file
// that was just created
sprintf(baseCommand, "sbatch %s\n", slurmFilename);
}
// else we're running an interactive job
else {
char iCom[1024];
int len;
len = 0;
// set the job name
len += sprintf(iCom+len, "--job-name=CHPL-%.10s ",basenamePtr);
// suppress informational messages, will still display errors
len += sprintf(iCom+len, "--quiet ");
// request the number of locales, with 1 task per node, and number of cores
// cpus-per-task. We probably don't need --nodes and --ntasks specified
// since 1 task-per-node with n --tasks implies -n nodes
len += sprintf(iCom+len, "--nodes=%d ",numLocales);
len += sprintf(iCom+len, "--ntasks=%d ", numLocales);
len += sprintf(iCom+len, "--ntasks-per-node=%d ", procsPerNode);
len += sprintf(iCom+len, "--cpus-per-task=%d ", getCoresPerLocale());
// request exclusive access
len += sprintf(iCom+len, "--exclusive ");
// kill the job if any program instance halts with non-zero exit status
len += sprintf(iCom+len, "--kill-on-bad-exit ");
// Set the walltime if it was specified
if (walltime) {
len += sprintf(iCom+len, "--time=%s ",walltime);
}
// Set the nodelist if it was specified
if (nodelist) {
len += sprintf(iCom+len, "--nodelist=%s ", nodelist);
}
// Set the partition if it was specified
if (partition) {
len += sprintf(iCom+len, "--partition=%s ", partition);
}
// Set the exclude list if it was specified
if (exclude) {
len += sprintf(iCom+len, "--exclude=%s ", exclude);
}
// set any constraints
if (constraint) {
len += sprintf(iCom+len, " --constraint=%s ", constraint);
}
// set the account name if one was provided
if (account && strlen(account) > 0) {
len += sprintf(iCom+len, "--account=%s ", account);
}
// add the (possibly wrapped) binary name
len += sprintf(iCom+len, "%s %s ",
chpl_get_real_binary_wrapper(), chpl_get_real_binary_name());
// add any arguments passed to the launcher to the binary
for (i=1; i<argc; i++) {
len += sprintf(iCom+len, "%s ", argv[i]);
}
// launch the job using srun
sprintf(baseCommand, "srun %s ", iCom);
}
// copy baseCommand into command and return it
size = strlen(baseCommand) + 1;
command = chpl_mem_allocMany(size, sizeof(char), CHPL_RT_MD_COMMAND_BUFFER, -1, 0);
sprintf(command, "%s", baseCommand);
if (strlen(command)+1 > size) {
chpl_internal_error("buffer overflow");
}
return command;
}
