[Question] expected throughput of cloud tpu on embedding lookup?
まだ誰も着手していません。
評価
- 難易度
- 4/5
- 見積もり時間
- 3〜5日
- 初心者へのやさしさ
- 12/100
- issue の種類
- ドキュメント
- 明瞭さ
- 説明が足りない
- 活発さ
- 停滞
- 技術スタック
- jupyter-notebook, python, tensorflow
調査の方向性
The benchmark script lives in the issue body and calls tpu_embedding_layer.TPUEmbedding from tensorflow_recommenders/layers/embedding/tpu_embedding_layer.py. The question is what throughput to expect and how to time TPU embedding lookups correctly, which needs TPU runtime knowledge rather than a code change. Done would be a reference figure or a corrected timing approach, and the issue has no comments or assignees to confirm either.
索引モデルが issue の本文から書いたものです。
説明
(duplicated question as in https://github.com/tensorflow/recommenders/issues/579, just not sure which repo is a better place for this kind of question)
Hi,
I read this blog recently https://cloud.google.com/blog/topics/developers-practitioners/building-large-scale-recommenders-using-cloud-tpus, very interested in it and wondering the raw performance of TPUEmbedding lookup performance.(we can quite easily get the perf data of tf.nn.(safe)embedding_lookup(_sparse) etc. but it becomes harder to get TPUEmbedding lookup perf data)
Based the test script included in this repo, I wrote piece of benchmarking code to test it:
# Copyright 2022 The TensorFlow Recommenders Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# https://github.com/tensorflow/recommenders/blob/main/tensorflow_recommenders/layers/embedding/tpu_embedding_layer_test.py
# trying to test tpu embedding lookup throughput, while I could not find lower-level API for doing such test
import time
import numpy as np
import tensorflow as tf
from tensorflow_recommenders.layers.embedding import tpu_embedding_layer
TABLE_SIZE = 1000000
EMB_DIM = 128
QUERY_KEY_NUM = 65536 * 8 # * 64, killed as Allocation of xxx exceeds 10% of free system memory.
class TPUEmbeddingLayerTest():
def __init__(self):
self.embedding_values = np.arange(TABLE_SIZE * EMB_DIM, dtype=np.float32)
self.initializer = tf.constant_initializer(self.embedding_values)
self.table_config = tf.tpu.experimental.embedding.TableConfig(
vocabulary_size=TABLE_SIZE,
dim=EMB_DIM,
initializer=self.initializer,
combiner='sum',
name='embedding_table')
self.feature_config = {
'indices2embeddings': tf.tpu.experimental.embedding.FeatureConfig(
table=self.table_config, name='indices2embeddings'),
}
self.batch_size = QUERY_KEY_NUM
self.sample_size = 1
# TODO(pehuang): draw samples randomly by given distribution
self.data_point_indices = np.zeros((self.batch_size, 2), dtype=np.int32)
self.data_point_indices[:, 0] = np.arange(self.batch_size, dtype=np.int32)
self.data_points = np.random.choice(TABLE_SIZE, QUERY_KEY_NUM)
self.embedding_lookup_input_data = tf.SparseTensor(
indices=self.data_point_indices,
values=tf.convert_to_tensor(self.data_points, dtype=tf.int32), # fp64 embedding and int32 key by default?
dense_shape=[self.batch_size, self.sample_size])
self.dataset = tf.data.Dataset.from_tensors({'indices2embeddings': self.embedding_lookup_input_data})
def embedding_lookup_throughput_test(self, optimizer_name='sgd', training=False):
# resolver = tf.distribute.cluster_resolver.TPUClusterResolver('').connect('')
# strategy = tf.distribute.TPUStrategy(resolver)
strategy = tf.distribute.get_strategy() # Use the default strategy.
with strategy.scope():
embedding_layer = tpu_embedding_layer.TPUEmbedding(feature_config=self.feature_config, optimizer=None)
input_args = {'batch_size': self.batch_size,
'shape': (),
'sparse': True,
'dtype': tf.int32}
inputs = {'indices2embeddings': tf.keras.Input(**input_args, name='indices2embeddings')}
embeddings = embedding_layer(inputs)
self.model = tf.keras.Model(inputs=(inputs), outputs=(embeddings))
dist = strategy.experimental_distribute_dataset(self.dataset, options=tf.distribute.InputOptions(experimental_fetch_to_device=False))
dist_iter = iter(dist)
def lookup(features):
res = self.model(features)
return res
# for _ in range(10): # warmup 10 rounds failed after batch data used up, and raised StopIteration exception
# result = strategy.run(lookup, args=(next(dist_iter),))
t_start = time.time()
result = strategy.run(lookup, args=(next(dist_iter),))
t_end = time.time()
# import pdb; pdb.set_trace() # stop to check embeddings
print("embedding throughput: {}GB/s".format((QUERY_KEY_NUM * EMB_DIM) / 1e9 / (t_end - t_start)))
# embedding throughput: 0.04273784880155698GB/s, must be cold-start, D2H and H2D etc. included
# but I cannot warmup/increase table size temporally, using cloud shell provided tpu
if __name__ == '__main__':
test = TPUEmbeddingLayerTest()
test.embedding_lookup_throughput_test()
I used cloud shell to get access to tpu resource as introduced in https://github.com/tensorflow/tpu, but the reported throughput data is quite below expectation(obviously, its due to various issues in my benchmark script), may anyone help correction the script to get benchmark results in the right way? Or just provide some reference data, so I could know what the expected throughput should be?
THX!
- 主要言語
- Jupyter Notebook
- スター
- 5.3k
- フォーク
- 1.7k
- PR マージ指標
- 30日以内にマージされた PR はありません
環境構築
このプロジェクトには開発コンテナ、Dockerfile、コントリビューションガイドがありません。まず README を読み、一般的な手順ははじめてのコントリビューションガイドを参照してください。
はじめの一歩
- issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
- 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
- リポジトリをフォークし、ブランチを切って変更します。
- issue 番号を参照したプルリクエストを送ります。
tensorflow/tpu のほかの issue
-
難易度 2/5 半日 初心者へのやさしさ 62/100
tensorflow/tpu#3542 ·
-
imagenet_to_gcs.pyオープン
難易度 1/5 1時間未満 初心者へのやさしさ 78/100
tensorflow/tpu#951 · リアクション 2 件 ·
-
Broken linkオープン
難易度 1/5 1時間未満 初心者へのやさしさ 70/100
tensorflow/tpu#908 ·
-
難易度 1/5 1時間未満 初心者へのやさしさ 62/100
tensorflow/tpu#889 · コメント 1 件 · リアクション 1 件 ·
-
難易度 1/5 1時間未満 初心者へのやさしさ 78/100
tensorflow/tpu#582 · コメント 1 件 ·
似ている issue
-
needs-triage
難易度 1/5 1時間未満 初心者へのやさしさ 88/100
flashinfer-ai/flashinfer#9803 ·
メンテナーはふだん 1 日以内に返信
-
bug module: rl
難易度 2/5 1〜3時間 初心者へのやさしさ 78/100
NVIDIA/Megatron-LM#9644 ·
メンテナーはふだん 2 日以内に返信
-
`build_flavor_config_from_local_checkpoint` fails with transformers 5.23.0: pipeline `impl` is now a dotted-path string対応中かも @mlflow-app が今日担当しました。 オープンhas-closing-pr nightly triage: reproduced
難易度 2/5 1〜3時間 初心者へのやさしさ 78/100
mlflow/mlflow#28000 · コメント 2 件 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 76/100
X-PLUG/MobileAgent#304 ·
-
難易度 2/5 1〜3時間 初心者へのやさしさ 78/100
FluxML/Optimisers.jl#254 ·