reward no longer terminal

Author: Graeme22

cluster.jl

@@ -11,7 +11,7 @@ using .job
 
 export ClusterEnv, QUEUE_SIZE, ZONES
 
-const QUEUE_SIZE, ZONES, SLICE_SIZE = 64, 16, 50_000
+const QUEUE_SIZE, ZONES, SLICE_SIZE = 64, 16, 8192
 
 struct Metrics
     avg_bounded_slowdown::Float32
@@ -212,6 +212,7 @@ function (env::ClusterEnv)(action)
             for i in 1:length(env.cluster)
                 if env.cluster[i].simulated_run_time >= env.cluster[i].run_time
                     env.available_cores += env.cluster[i].cores
+                    env.reward += env.sjf_bsld - bounded_slowdown(env.cluster[i])
                     push!(to_remove, i)
                 end
             end
@@ -248,7 +249,7 @@ function (env::ClusterEnv)(action)
             # currently: negative average bounded slowdown relative to SJF
             # positive reward = better performance than SJF;
             # negative = worse; zero = same
-            env.reward = env.workload.sjf_bsld - avg_bsld
+            #env.reward = env.workload.sjf_bsld - avg_bsld
             env.done = true
             env.metrics = Metrics(avg_bsld, avg_wait_time, max_wait_time)#, avg_utilization)
             break

run-cpu.jl

@@ -12,15 +12,15 @@ using .cluster
 
 save_dir = "save"
 N_ENV = 8
-TRAJECTORY_SIZE = 1024
-N_TRAIN_ITERATION = 50
-N_EPOCH = 500
+TRAJECTORY_SIZE = 64
+N_TRAIN_ITERATION = 150
+N_EPOCH = 50
 env = MultiThreadEnv([ClusterEnv() for i in 1:N_ENV])
 
 shared_network = Chain(
     x -> reshape(x, ZONES + 4, QUEUE_SIZE, :),
-    Dense(ZONES + 4, 512, relu),
-    Dense(512, 1),
+    Dense(ZONES + 4, 128, relu),
+    Dense(128, 1),
     x -> reshape(x, QUEUE_SIZE, :)  # squeeze
 )
 
@@ -30,8 +30,8 @@ agent = Agent(
             actor = shared_network,
             critic = Chain(
                 shared_network,
-                Dense(QUEUE_SIZE, 256, relu),
-                Dense(256, 1),
+                Dense(QUEUE_SIZE, 128, relu),
+                Dense(128, 1),
                 x -> reshape(x, :)  # squeeze
             ),
             optimizer = Adam(1e-5)