tensorflow/core/profiler/g3doc/profile_time.md
When we run a model, Tensorflow schedules and runs the nodes (operations) in the graph. An operation can be placed on an accelerator or on CPU.
When an operation is placed on accelerator, it will first be scheduled by TensorFlow on CPU. Normally, it's the code in OpKernel::Compute. OpKernel::Compute can decide to dispatch some of the computations on the accelerator. While some computation (e.g. pre-processing) is still done in CPU. OpKernel::Compute can dispatch computation on accelerator and return, or it can also wait for the accelerator to finish.
tfprof reports 3 execution times:
Since accelerator, such as GPU, usually runs operation asynchronously, you might notice an operation finishes on cpu before it starts running on accelerator.
When an operation is placed on CPU, it will completely run on CPU. Hence, <b>exec_micros</b> is equal to <b>cpu_micros</b> and <b>accelerator_micros</b> should be 0.
# In code view, the time of each line of Python code is the aggregated
# times of all operations created by that line.
# In command line, it requires --graph_path --op_log_path and --run_meta_path.
# --op_log_path provides the code traces information.
# --run_meta_path provides the time information.
tfprof> code -show_name_regexes seq2seq_attention.* -max_depth 10 -select micros -order_by micros
node name | execution time
_TFProfRoot (--/3.74sec)
seq2seq_attention.py'>:168:run_filename_from...:none (0us/3.74sec)
seq2seq_attention.py'>:33:_run_code_in_main:none (0us/3.74sec)
seq2seq_attention.py:316:<module>:app.run() (0us/3.74sec)
seq2seq_attention.py:270:main:_Train(model, bat... (0us/3.74sec)
seq2seq_attention.py:128:_Train:model.build_graph() (0us/3.74sec)
seq2seq_attention_model.py:360:build_graph:self._add_seq2seq() (0us/2.79sec)
seq2seq_attention_model.py:293:_add_seq2seq:decoder_outputs, ... (0us/2.46sec)
seq2seq_attention_model.py:192:_add_seq2seq:sequence_length=a... (0us/265.31ms)
seq2seq_attention_model.py:253:_add_seq2seq:initial_state_att... (0us/50.35ms)
seq2seq_attention_model.py:173:_add_seq2seq:for x in encoder_... (0us/8.72ms)
seq2seq_attention_model.py:218:_add_seq2seq:w_t = tf.transpos... (0us/2.39ms)
...
seq2seq_attention_model.py:363:build_graph:self._add_train_o... (0us/949.10ms)
seq2seq_attention_model.py:307:_add_train_op:tf.gradients(self... (0us/641.44ms)
seq2seq_attention_model.py:322:_add_train_op:zip(grads, tvars)... (0us/307.56ms)
...
seq2seq_attention_model.py:364:build_graph:self._summaries =... (0us/13us)
seq2seq_attention_model.py:361:build_graph:self.global_step ... (0us/12us)
...
seq2seq_attention.py:129:_Train:saver = tf.train.... (0us/0us)
seq2seq_attention.py:140:_Train:global_step=model... (0us/0us)
# Sometimes you want to explore a specific function. You can do that
# with -start_name_regexes.
tfprof> code -start_name_regexes .*_add_seq2seq.* -show_name_regexes seq2seq_attention.* -max_depth 10 -select micros -order_by micros
node name | execution time
_TFProfRoot (--/3.74sec)
seq2seq_attention_model.py:360:build_graph:self._add_seq2seq() (0us/2.79sec)
seq2seq_attention_model.py:293:_add_seq2seq:decoder_outputs, ... (0us/2.46sec)
seq2seq_attention_model.py:289:sampled_loss_func:num_classes=vsize) (0us/2.46sec)
seq2seq_attention_model.py:282:sampled_loss_func:labels = tf.resha... (0us/164us)
# You can also dive deeper into tensorflow's libraries.
tfprof> code -max_depth 5 -select micros -order_by micros -start_name_regexes .*_add_seq2seq.* -min_micros 100000
_TFProfRoot (--/3.74sec)
seq2seq_attention_model.py:360:build_graph:self._add_seq2seq() (0us/2.79sec)
seq2seq_attention_model.py:293:_add_seq2seq:decoder_outputs, ... (0us/2.46sec)
seq2seq_lib.py:181:sampled_sequence_...:average_across_ti... (0us/2.46sec)
seq2seq_lib.py:147:sequence_loss_by_...:crossent = loss_f... (0us/2.46sec)
seq2seq_attention_model.py:192:_add_seq2seq:sequence_length=a... (0us/265.31ms)
seq2seq_lib.py:104:bidirectional_rnn:sequence_length, ... (0us/127.27ms)
core_rnn.py:195:static_rnn:state_size=cell.s... (0us/127.20ms)
seq2seq_lib.py:110:bidirectional_rnn:initial_state_bw,... (0us/125.96ms)
core_rnn.py:195:static_rnn:state_size=cell.s... (0us/125.86ms)
# It can also be done in Python API
opts = model_analyzer.TRAINABLE_VARS_PARAMS_STAT_OPTIONS.copy()
opts['account_type_regexes'] = ['.*']
opts['show_name_regexes'] = ['.*model_analyzer_testlib.py.*']
opts['account_displayed_op_only'] = False
opts['select'] = ['micros']
tfprof_node = model_analyzer.print_model_analysis(
sess.graph, run_meta, cmd='code', options=opts)
You can generate some visualization in code view:
Set -output timeline:outfile=<filename> to generate timeline instead of stdout.
<left>
</left>
# In op view, you can view the aggregated time of each operation type.
tfprof> op -select micros,occurrence -order_by micros
node name | execution time | op occurrence
SoftmaxCrossEntropyWithLogits 1.37sec (100.00%, 36.44%), 30
MatMul 618.97ms (63.56%, 16.51%), 3450
Add 273.76ms (47.06%, 7.30%), 2180
Sub 215.41ms (39.76%, 5.74%), 4372
ConcatV2 203.88ms (34.01%, 5.44%), 6098
Mul 134.32ms (28.58%, 3.58%), 9427
ApplyAdam 92.66ms (25.00%, 2.47%), 27
Switch 72.43ms (22.53%, 1.93%), 30654
LogUniformCandidateSampler 69.01ms (20.59%, 1.84%), 30
Unique 53.50ms (18.75%, 1.43%), 2
AddN 50.10ms (17.33%, 1.34%), 5481
# You might be surprised to see that SoftmaxCrossEntropyWithLogits is
# that expensive. As shown below, it is placed on cpu.
tfprof> op -select micros,device -order_by micros
node name | execution time | assigned devices
SoftmaxCrossEntropyWithLogits 1.37sec (100.00%, 36.44%), /job:worker/replica:0/task:0/cpu:0
MatMul 618.97ms (63.56%, 16.51%), |/job:worker/replica:0/task:0/cpu:0|/job:worker/replica:0/task:0/device:GPU:0|/job:worker/replica:0/task:0/device:GPU:1|/job:worker/replica:0/task:0/device:GPU:2|/job:worker/replica:0/task:0/device:GPU:3
Usually, use graph view to generate a timeline to visualize the result.
In the chrome://tracing UI, click "Flow Event" in "View Options" of upper right corner to see the flow of tensors.
<left> TODO(xpan): Show the image correctly in github.  </left>tfprof options allow users to generate timeline in some advanced ways.
# Only generate timeline for gpu3 and cpu on workers.
graph -max_depth 10000000 -step 0 -account_type_regexes .*gpu:3.*,.*worker.*cpu:0.* -output timeline:outfile=<filename.json>
generating trace file.
******************************************************
Timeline file is written to <filename.json>.
Open a Chrome browser, enter URL chrome://tracing and load the timeline file.
******************************************************
Usually scope view allows you to pin point the problematic places if you have properly named your operations with tf.name_scope or tf.variable_scope.
tfprof> scope -max_depth 30 -select micros -min_micros 100000 -order_by micros
node name | execution time
_TFProfRoot (--/8.12sec)
tower_3/gradients/tower_3/Conv2d_1a_3x3/convolution_grad/Conv2DBackpropFilter (126.34ms/126.34ms)
tower_1/gradients/tower_1/Conv2d_1a_3x3/convolution_grad/Conv2DBackpropFilter (125.44ms/125.44ms)
tower_2/gradients/tower_2/Conv2d_1a_3x3/convolution_grad/Conv2DBackpropFilter (124.85ms/124.85ms)
tower_0/gradients/tower_0/Conv2d_1a_3x3/convolution_grad/Conv2DBackpropFilter (124.45ms/124.45ms)