update

src/evaluator.go

@@ -66,11 +66,6 @@ func (eva *Evaluator) add(node NodeStatus, task Task) {
 		eva.totalWorker++
 	}
 
-	if task.IsPS {
-		//eva.costLoad += 1
-	} else {
-		//eva.costLoad += 0.5
-	}
 	numberGPU := 1
 	for _, gpu := range node.Status {
 		if gpu.MemoryAllocated != 0 {
@@ -104,11 +99,6 @@ func (eva *Evaluator) remove(node NodeStatus, task Task) {
 		eva.totalWorker--
 	}
 
-	if task.IsPS {
-		//eva.costLoad -= 1
-	} else {
-		//eva.costLoad -= 0.5
-	}
 	numberGPU := 1
 	for _, gpu := range node.Status {
 		if gpu.MemoryAllocated != 0 {

src/job_manager.go

@@ -80,7 +80,7 @@ func (jm *JobManager) start() {
 				v.Set("mem_limit", strconv.Itoa(jm.job.Tasks[index].Memory)+"m")
 				v.Set("cpu_limit", strconv.Itoa(jm.job.Tasks[index].NumberCPU))
 				v.Set("network", jm.network)
-				v.Set("should_wait", "1")
+				v.Set("should_wait", "0")
 				v.Set("output_dir", "/tmp/")
 				v.Set("hdfs_address", "http://192.168.100.104:50070/")
 				v.Set("hdfs_dir", "/user/yao/output/"+jm.job.Name)

src/resource_pool.go

@@ -896,7 +896,12 @@ func (pool *ResourcePool) releaseResource(job Job, agent NodeStatus) {
 	seg.Lock.Lock()
 	defer seg.Lock.Unlock()
 
-	node := seg.Nodes[agent.ClientID]
+	node, ok := seg.Nodes[agent.ClientID]
+	if !ok {
+		/* in case node is offline */
+		/* TODO, update usingTotalGPU correctly */
+		return
+	}
 	for _, gpu := range agent.Status {
 		for j := range node.Status {
 			if gpu.UUID == node.Status[j].UUID {