From 8d009ce3b3a511289ea641ee917cb1e90ddd82aa Mon Sep 17 00:00:00 2001
From: David Anderson <davea@ssl.berkeley.edu>
Date: Sat, 24 May 2014 15:18:41 -0700
Subject: [PATCH] client: scheduling and work fetch tweaks for GPU exclusion
 cases

Scheduling: if a resource has exclusions, put all jobs in the run list;
otherwise we might fail to have a job for a GPU instance, and starve it.

Work fetch: allow work fetch from zero-share projects if the resource
has instances that are idle because of GPU exclusion
---
 client/cpu_sched.cpp  | 16 ++++++----------
 client/work_fetch.cpp |  6 ++++--
 2 files changed, 10 insertions(+), 12 deletions(-)

diff --git a/client/cpu_sched.cpp b/client/cpu_sched.cpp
index 29eb03b8a9..584ad4d6a9 100644
--- a/client/cpu_sched.cpp
+++ b/client/cpu_sched.cpp
@@ -170,23 +170,19 @@ struct PROC_RESOURCES {
     void schedule(RESULT* rp, ACTIVE_TASK* atp, const char* description) {
         if (log_flags.cpu_sched_debug) {
             msg_printf(rp->project, MSG_INFO,
-                "[cpu_sched_debug] scheduling %s (%s) (prio %f)",
+                "[cpu_sched_debug] add to run list: %s (%s) (prio %f)",
                 rp->name, description,
                 rp->project->sched_priority
             );
         }
         if (rp->uses_coprocs()) {
-            // if this job is currently running,
-            // and the resource type has exclusions,
-            // don't reserve instances;
-            // This allows more jobs in the run list
-            // and avoids a starvation case
+            // if the resource type has exclusions, don't reserve instances.
+            // It means that the run list will include all jobs
+            // for that resource type.
+            // Inefficient, but necessary to avoid starvation cases.
             //
             int rt = rp->avp->gpu_usage.rsc_type;
-            bool dont_reserve =
-                rsc_work_fetch[rt].has_exclusions
-                && atp != NULL
-                && atp->is_gpu_task_running();
+            bool dont_reserve = rsc_work_fetch[rt].has_exclusions;
             if (!dont_reserve) {
                 reserve_coprocs(*rp);
             }
diff --git a/client/work_fetch.cpp b/client/work_fetch.cpp
index acb0575a23..ee7ed0624e 100644
--- a/client/work_fetch.cpp
+++ b/client/work_fetch.cpp
@@ -270,7 +270,9 @@ void RSC_WORK_FETCH::print_state(const char* name) {
         shortfall, nidle_now, saturated_time,
         busy_time_estimator.get_busy_time()
     );
-    //msg_printf(0, MSG_INFO, "[work_fetch] sim used inst %d sim excl inst %d", sim_used_instances, sim_excluded_instances);
+    msg_printf(0, MSG_INFO, "[work_fetch] sim used inst %d sim excl inst %d",
+        sim_used_instances, sim_excluded_instances
+    );
     for (unsigned int i=0; i<gstate.projects.size(); i++) {
         char buf[256];
         PROJECT* p = gstate.projects[i];
@@ -603,7 +605,7 @@ int RSC_WORK_FETCH::cant_fetch(PROJECT *p) {
     // if project has zero resource share,
     // only fetch work if a device is idle
     //
-    if (p->resource_share == 0 && nidle_now == 0) {
+    if (p->resource_share == 0 && nidle_now == 0 && sim_excluded_instances==0) {
         WF_DEBUG(msg_printf(p, MSG_INFO, "skip: zero share");)
         return DONT_FETCH_ZERO_SHARE;
     }