1_machine_4_gpu.benchmark

benchmark/alexnet/1_machine_2_gpu.placement

 placement_group {
   op_set {
     op_name: "decode"
-    op_name: "transpose"
   }
   parallel_conf {
     policy: kDataParallel
@@ -11,6 +10,7 @@ placement_group {
 
 placement_group {
   op_set {
+    op_name: "transpose"
     op_name: "conv1"
     op_name: "pool1"
     op_name: "conv2"

benchmark/alexnet/1_machine_4_gpu.job

+net: "./train.net"
+resource: "./1_machine_4_gpu.resource"
+placement: "./1_machine_4_gpu.placement"
+other: "./piece_size_1024.other"

benchmark/alexnet/1_machine_4_gpu.placement

+placement_group {
+  op_set {
+    op_name: "decode"
+  }
+  parallel_conf {
+    policy: kDataParallel
+    device_name: "first:cpu:32"
+  }
+}
+
+placement_group {
+  op_set {
+    op_name: "transpose"
+    op_name: "conv1"
+    op_name: "pool1"
+    op_name: "conv2"
+    op_name: "pool2"
+    op_name: "conv3"
+    op_name: "conv4"
+    op_name: "conv5"
+    op_name: "pool5"
+    op_name: "dropout1"
+    op_name: "dropout2"
+    op_name: "softmax_loss"
+  }
+  parallel_conf {
+    policy: kDataParallel
+    device_name: "first:gpu:0-3"
+  }
+}
+
+placement_group {
+  op_set {
+    op_name: "fc1"
+    op_name: "fc2"
+    op_name: "fc3"
+  }
+  parallel_conf {
+    policy: kDataParallel
+    device_name: "first:gpu:0-3"
+  }
+}

benchmark/alexnet/report.md

 batch_size: 1024
 
-gpu num | time per batch
+gpu num | time (one batch)
 :-------| :-------------
-1       | 534ms
+1       | 551ms
+2       | 282ms
+4       | 269ms