sched: Add GPU scheduling

Merge @TWRS' change and slightly modify it.

Propagte the gpu request information received
from flux submit to the request input object
for scheduling.

gpu becomes a constraint for resrc's resource
type matching logic.

sched/sched.c

@@ -422,6 +422,7 @@ static inline int fill_resource_req (flux_t *h, flux_lwj_t *j, json_t *jcb)
     int rc = -1;
     int64_t nn = 0;
     int64_t nc = 0;
+    int64_t ngpus = 0;
     int64_t walltime = 0;
     json_t *o = NULL;
     ssrvctx_t *ctx = getctx (h);
@@ -437,9 +438,12 @@ static inline int fill_resource_req (flux_t *h, flux_lwj_t *j, json_t *jcb)
         goto done;
     if (!Jget_int64 (o, JSC_RDESC_NCORES, &nc))
         goto done;
+    if (!Jget_int64 (o, JSC_RDESC_NGPUS, &ngpus))
+        goto done;
 
     j->req->nnodes = (uint64_t) nn;
     j->req->ncores = (uint64_t) nc;
+    j->req->ngpus = (uint64_t) ngpus;
     if (!Jget_int64 (o, JSC_RDESC_WALLTIME, &walltime) || !walltime) {
         j->req->walltime = (uint64_t) 3600;
     } else {
@@ -1473,8 +1477,6 @@ static resrc_reqst_t *get_resrc_reqst (ssrvctx_t *ctx, flux_lwj_t *job,
      */
     req_res = Jnew ();
     if (job->req->nnodes > 0) {
-        json_t *child_core = Jnew ();
-
         Jadd_str (req_res, "type", "node");
         Jadd_int64 (req_res, "req_qty", job->req->nnodes);
         *nreqrd = job->req->nnodes;
@@ -1493,6 +1495,7 @@ static resrc_reqst_t *get_resrc_reqst (ssrvctx_t *ctx, flux_lwj_t *job,
             Jadd_bool (req_res, "exclusive", false);
         }
 
+        json_t *child_core = Jnew ();
         Jadd_str (child_core, "type", "core");
         Jadd_int64 (child_core, "req_qty", job->req->corespernode);
         /* setting size == 1 devotes (all of) the core to the job */
@@ -1501,7 +1504,22 @@ static resrc_reqst_t *get_resrc_reqst (ssrvctx_t *ctx, flux_lwj_t *job,
         Jadd_bool (child_core, "exclusive", true);
         Jadd_int64 (child_core, "starttime", starttime);
         Jadd_int64 (child_core, "endtime", starttime + job->req->walltime);
-        json_object_set_new (req_res, "req_child", child_core);
+
+        json_t *children = Jnew_ar();
+        json_array_append_new (children, child_core);
+        if (job->req->ngpus) {
+            json_t *child_gpu = Jnew ();
+            Jadd_str (child_gpu, "type", "gpu");
+            Jadd_int64 (child_gpu, "req_qty", job->req->ngpus);
+            /* setting size == 1 devotes (all of) the gpu to the job */
+            Jadd_int64 (child_gpu, "req_size", 1);
+            /* setting exclusive to true prevents multiple jobs per core */
+            Jadd_bool (child_gpu, "exclusive", true);
+            Jadd_int64 (child_gpu, "starttime", starttime);
+            Jadd_int64 (child_gpu, "endtime", starttime + job->req->walltime);
+            json_array_append_new (children, child_gpu);
+        }
+        json_object_set_new (req_res, "req_children", children);
     } else if (job->req->ncores > 0) {
         Jadd_str (req_res, "type", "core");
         Jadd_int (req_res, "req_qty", job->req->ncores);

sched/scheduler.h

@@ -43,6 +43,7 @@
 typedef struct flux_resources {
     uint64_t nnodes; /*!< num of nodes requested by a job */
     uint64_t ncores; /*!< num of cores requested by a job */
+    uint64_t ngpus;  /*!< num of gpus requested by a job */
     uint64_t corespernode; /*!< num of cores per node requested by a job */
     uint64_t walltime; /*!< walltime requested by a job */
     bool     node_exclusive; /*!< job requires exclusive use of node if true */