1## Auto Detect and Advise 2 3tfprof analyzes profiles and generates advice for common issues. 4 5### Run Advise. 6 7```python 8# First create a profiler. See profiler tutorials for more details. 9profiler = tf.profiler.Profiler(sess.graph) 10run_meta = config_pb2.RunMetadata() 11_ = sess.run(r1, 12 options=config_pb2.RunOptions( 13 trace_level=config_pb2.RunOptions.FULL_TRACE), 14 run_metadata=run_meta) 15profiler.add_step(1, run_meta) 16 17# Then Start advise. 18profiler.advise() 19 20# For one-shot API 21tf.profiler.advise( 22 sess.graph, run_meta=run_metadata) 23``` 24 25```shell 26# Run advisor on CLI 27# See CLI tutorial on generating the files. 28tfprof --graph_path=graph.pbtxt \ 29 --run_meta_path=run_metadata \ 30 --op_log_path=tfprof_log 31 32tfprof> advise 33AcceleratorUtilizationChecker: 34device: /job:worker/replica:0/task:0/device:GPU:0 low utilization: 0.03 35device: /job:worker/replica:0/task:0/device:GPU:1 low utilization: 0.08 36device: /job:worker/replica:0/task:0/device:GPU:2 low utilization: 0.04 37device: /job:worker/replica:0/task:0/device:GPU:3 low utilization: 0.21 38 39OperationChecker: 40Found operation using NHWC data_format on GPU. Maybe NCHW is faster. 41 42ExpensiveOperationChecker: 43top 1 operation type: SoftmaxCrossEntropyWithLogits, cpu: 1.37sec, accelerator: 0us, total: 1.37sec (26.68%) 44top 2 operation type: MatMul, cpu: 427.39ms, accelerator: 280.76ms, total: 708.14ms (13.83%) 45top 3 operation type: ConcatV2, cpu: 357.83ms, accelerator: 31.80ms, total: 389.63ms (7.61%) 46seq2seq_attention_model.py:360:build_graph:self._add_seq2seq(), cpu: 3.16sec, accelerator: 214.84ms, total: 3.37sec 47 seq2seq_attention_model.py:293:_add_seq2seq:decoder_outputs, ..., cpu: 2.46sec, accelerator: 3.25ms, total: 2.47sec 48 seq2seq_lib.py:181:sampled_sequence_...:average_across_ti..., cpu: 2.46sec, accelerator: 3.24ms, total: 2.47sec 49 seq2seq_lib.py:147:sequence_loss_by_...:crossent = loss_f..., cpu: 2.46sec, accelerator: 3.06ms, total: 2.46sec 50 seq2seq_attention_model.py:289:sampled_loss_func:num_classes=vsize), cpu: 2.46sec, accelerator: 3.06ms, total: 2.46sec 51 seq2seq_attention_model.py:282:sampled_loss_func:labels = tf.resha..., cpu: 164us, accelerator: 0us, total: 164us 52 seq2seq_lib.py:148:sequence_loss_by_...:log_perp_list.app..., cpu: 1.33ms, accelerator: 120us, total: 1.45ms 53 seq2seq_lib.py:151:sequence_loss_by_...:total_size = tf.a..., cpu: 154us, accelerator: 23us, total: 177us 54 seq2seq_lib.py:184:sampled_sequence_...:return cost / tf...., cpu: 97us, accelerator: 8us, total: 105us 55 math_ops.py:690:cast:return gen_math_o..., cpu: 62us, accelerator: 3us, total: 65us 56 math_ops.py:839:binary_op_wrapper:return func(x, y,..., cpu: 35us, accelerator: 5us, total: 40us 57 seq2seq_attention_model.py:192:_add_seq2seq:sequence_length=a..., cpu: 651.56ms, accelerator: 158.92ms, total: 810.48ms 58 seq2seq_lib.py:104:bidirectional_rnn:sequence_length, ..., cpu: 306.58ms, accelerator: 73.54ms, total: 380.12ms 59 core_rnn.py:195:static_rnn:state_size=cell.s..., cpu: 306.52ms, accelerator: 73.54ms, total: 380.05ms 60 rnn.py:218:_rnn_step:_maybe_copy_some_..., cpu: 303.76ms, accelerator: 73.54ms, total: 377.30ms 61 rnn.py:216:_rnn_step:time >= max_seque..., cpu: 2.75ms, accelerator: 0us, total: 2.75ms 62 core_rnn.py:179:static_rnn:max_sequence_leng..., cpu: 67us, accelerator: 0us, total: 67us 63 seq2seq_lib.py:110:bidirectional_rnn:initial_state_bw,..., cpu: 296.21ms, accelerator: 73.54ms, total: 369.75ms 64 core_rnn.py:195:static_rnn:state_size=cell.s..., cpu: 296.11ms, accelerator: 73.54ms, total: 369.65ms 65 rnn.py:218:_rnn_step:_maybe_copy_some_..., cpu: 292.04ms, accelerator: 73.54ms, total: 365.58ms 66 rnn.py:216:_rnn_step:time >= max_seque..., cpu: 4.07ms, accelerator: 0us, total: 4.07ms 67 core_rnn.py:178:static_rnn:min_sequence_leng..., cpu: 85us, accelerator: 0us, total: 85us 68 core_rnn.py:179:static_rnn:max_sequence_leng..., cpu: 16us, accelerator: 0us, total: 16us 69 seq2seq_lib.py:113:bidirectional_rnn:outputs = [tf.con..., cpu: 46.88ms, accelerator: 3.87ms, total: 50.75ms 70 ...(omitted) 71top 1 graph node: seq2seq/loss/sampled_sequence_loss/sequence_loss_by_example/SoftmaxCrossEntropyWithLogits_11, cpu: 89.92ms, accelerator: 0us, total: 89.92ms 72top 2 graph node: train_step/update_seq2seq/output_projection/w/ApplyAdam, cpu: 84.52ms, accelerator: 0us, total: 84.52ms 73top 3 graph node: seq2seq/loss/sampled_sequence_loss/sequence_loss_by_example/SoftmaxCrossEntropyWithLogits_19, cpu: 73.02ms, accelerator: 0us, total: 73.02ms 74``` 75 76### Checker 77 78There is no magic behind advise mode. tfprof builds the profiles first, then 79it runs through a list of `Checkers`, each one responsible for checking one 80area with the profile and report issues. A `Checker` is like a plugin. 81 82For example: 83 84#### JobChecker (Not Available OSS) 85 86* Checks RecvTensor RPC latency and bandwidth. 87* Checks CPU/Memory utilization of the job. 88 89#### AcceleratorUtilization Checker 90* Checks what percentage of time the accelerator spends on computation. 91 92#### OperationChecker 93 94* Checks whether the operation runs with optimal options. 95* Checks if there is a better implementation to replace the current operation. 96 97#### ExpensiveOperationChecker 98 99* Checks the most expensive operation type. 100* Checks the most expensive graph nodes. 101* Checks the most expensive graph-building Python codes. 102 103#### Contribute Your Checker 104 105Follow examples of accelerator_utilization_checker.h 106 107 108 109