Rename opt.cpus to opt.cpus_per_task, distributed tasks appropriately

for various values of cpus_per_task.

src/srun/launch.c

@@ -109,7 +109,8 @@ launch(void *arg)
 	taskid = 0;
 	if (opt.distribution == SRUN_DIST_BLOCK) {
 		for (i=0; ((i<job->nhosts) && (taskid<opt.nprocs)); i++) {
-			for (j=0; ((j<job->cpus[i]) && (taskid<opt.nprocs)); j++) {
+			for (j=0; (((j*opt.cpus_per_task)<job->cpus[i]) && 
+			           (taskid<opt.nprocs)); j++) {
 				task_ids[i][j] = taskid++;
 				job->ntask[i]++;
 			}

src/srun/opt.c

@@ -151,7 +151,7 @@ struct poptOption runTable[] = {
 	{"ntasks", 'n', POPT_ARG_INT, &opt.nprocs, OPT_NPROCS,
 	 "number of tasks to run",
 	 "ntasks"},
-	{"cpus-per-task", 'c', POPT_ARG_INT, &opt.cpus, OPT_CPUS,
+	{"cpus-per-task", 'c', POPT_ARG_INT, &opt.cpus_per_task, OPT_CPUS,
 	 "number of cpus required per task",
 	 "ncpus"},
 	{"nodes", 'N', POPT_ARG_INT, &opt.nodes, OPT_NODES,
@@ -252,7 +252,7 @@ typedef struct env_vars {
 env_vars_t env_vars[] = {
 	{"SLURM_DEBUG", OPT_DEBUG, NULL},
 	{"SLURM_NPROCS", OPT_INT, &opt.nprocs},
-	{"SLURM_CPUS_PER_TASK", OPT_INT, &opt.cpus},
+	{"SLURM_CPUS_PER_TASK", OPT_INT, &opt.cpus_per_task},
 	{"SLURM_NNODES", OPT_INT, &opt.nodes},
 	{"SLURM_PARTITION", OPT_STRING, &opt.partition},
 	{"SLURM_STDINMODE", OPT_INPUT, &opt.input},
@@ -311,14 +311,6 @@ int initialize_and_process_args(int argc, char *argv[])
 	/* initialize options with argv */
 	opt_args(argc, argv);
 
-	/* validate results */
-	if (opt.distribution == SRUN_DIST_UNKNOWN) {
-		if (opt.nprocs <= opt.nodes)
-			opt.distribution = SRUN_DIST_CYCLIC;
-		else
-			opt.distribution = SRUN_DIST_BLOCK;
-	}
-
 #if	__DEBUG
 	opt_list();
 #endif
@@ -521,7 +513,7 @@ static void opt_default()
 	opt.progname = NULL;
 
 	opt.nprocs = 1;
-	opt.cpus = 1;
+	opt.cpus_per_task = 1;
 	opt.nodes = 0; /* nodes need not be set */
 	opt.time_limit = -1;
 	opt.partition = NULL;
@@ -836,7 +828,6 @@ static void opt_args(int ac, char **av)
 		} 
 	}
 
-	
 	if (!opt_verify(optctx, nnodes_set, cpus_set, nprocs_set)) {
 		poptPrintUsage(optctx, stderr, 0);
 		exit(1);
@@ -861,6 +852,16 @@ opt_verify(poptContext optctx,
 		verified = false;
 	}
 
+	if (opt.distribution == SRUN_DIST_UNKNOWN) {
+		if (opt.nprocs <= opt.nodes)
+			opt.distribution = SRUN_DIST_CYCLIC;
+		else
+			opt.distribution = SRUN_DIST_BLOCK;
+	}
+
+	if (opt.mincpus < opt.cpus_per_task)
+		opt.mincpus = opt.cpus_per_task;
+
 	if ((opt.job_name == NULL) && (remote_argc > 0))
 		opt.job_name = base_name(remote_argv[0]);
 
@@ -908,9 +909,9 @@ opt_verify(poptContext optctx,
 			verified = false;
 		}
 
-		if (opt.cpus <= 0) {
-			error("%s: invalid number of cpus per process (-n %d)\n",
-				opt.progname, opt.cpus);
+		if (opt.cpus_per_task <= 0) {
+			error("%s: invalid number of cpus per task (-c %d)\n",
+				opt.progname, opt.cpus_per_task);
 			verified = false;
 		}
 

src/srun/opt.h

@@ -92,9 +92,9 @@ typedef struct srun_options {
 	char *cwd;		/* current working directory	*/
 
 	int  nprocs;		/* --nprocs=n,      -n n	*/
-	int  cpus;		/* --cpus_per_task=n, -c n	*/
+	int  cpus_per_task;	/* --cpus-per-task=n, -c n	*/
 	int  nodes;		/* --nodes=n,       -N n	*/ 
-	int  time_limit;	/* --time,   -t		*/
+	int  time_limit;	/* --time,   -t			*/
 	char *partition;	/* --partition=n,   -p n   	*/
 	enum distribution_t
 		distribution;	/* --distribution=, -m dist	*/
@@ -113,7 +113,7 @@ typedef struct srun_options {
 	char *attach;		/* --attach=id	    -a id	*/ 
 	bool join;		/* --join, 	    -j		*/
 
-	/* no longer need these, they are set globally : */
+	/* no longer need these, they are set globally : 	*/
 	/*int verbose;*/	/* -v, --verbose		*/	
 	/*int debug;*/		/* -d, --debug			*/
 

src/srun/srun.c

@@ -298,7 +298,7 @@ allocate_nodes(void)
 
 	job.req_nodes      = opt.nodelist;
 
-	job.num_procs      = opt.nprocs;
+	job.num_procs      = opt.nprocs * opt.cpus_per_task;
 
 	if (opt.nodes > -1)
 		job.num_nodes = opt.nodes;
@@ -409,7 +409,7 @@ create_job_step(job_t *job)
 	req.job_id     = job->jobid;
 	req.user_id    = opt.uid;
 	req.node_count = job->nhosts;
-	req.cpu_count  = opt.nprocs;
+	req.cpu_count  = opt.nprocs * opt.cpus_per_task;
 	req.node_list  = job->nodelist;
 	req.relative   = false;
 	if (opt.distribution == SRUN_DIST_BLOCK)
@@ -594,7 +594,7 @@ run_batch_job(void)
 
 	job.req_nodes      = opt.nodelist;
 
-	job.num_procs      = opt.nprocs;
+	job.num_procs      = opt.nprocs * opt.cpus_per_task;
 
 	if (opt.nodes > -1)
 		job.num_nodes = opt.nodes;