[Question] expected throughput of cloud tpu on embedding lookup?
还没有人认领这个 Issue。
评估
- 难度
- 4/5
- 预计耗时
- 3-5 天
- 新手友好度
- 12/100
- Issue 类型
- 文档
- 描述清晰度
- 需要澄清
- 活跃度
- 停滞
- 技术栈
- jupyter-notebook, python, tensorflow
调研方向
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.
由索引模型根据 Issue 内容生成。
描述
(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!
- 主要语言
- Jupyter Notebook
- 星标
- 5.3k
- 派生
- 1.7k
- PR 合并指标
- 30 天内没有已合并 PR
环境准备
这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
tensorflow/tpu 的其他 Issue
-
难度 2/5 半天 新手友好度 62/100
tensorflow/tpu#3542 ·
-
难度 1/5 1 小时以内 新手友好度 78/100
tensorflow/tpu#951 · 2 个 reaction ·
-
Broken link未关闭
难度 1/5 1 小时以内 新手友好度 70/100
tensorflow/tpu#908 ·
-
难度 1/5 1 小时以内 新手友好度 62/100
tensorflow/tpu#889 · 1 条评论 · 1 个 reaction ·
-
难度 1/5 1 小时以内 新手友好度 78/100
tensorflow/tpu#582 · 1 条评论 ·
相似的 Issue
-
needs-triage
难度 1/5 1 小时以内 新手友好度 88/100
flashinfer-ai/flashinfer#9803 ·
维护者通常 1 天内回复
-
bug module: rl
难度 2/5 1-3 小时 新手友好度 78/100
NVIDIA/Megatron-LM#9644 ·
维护者通常 2 天内回复
-
`build_flavor_config_from_local_checkpoint` fails with transformers 5.23.0: pipeline `impl` is now a dotted-path string可能已有人在做 @mlflow-app 今天认领。 未关闭has-closing-pr nightly triage: reproduced
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 76/100
X-PLUG/MobileAgent#304 ·
-
难度 2/5 1-3 小时 新手友好度 78/100
FluxML/Optimisers.jl#254 ·