1. deploy worker, parameter server on kubernetes cluster
1.1 build container image of worker, parameter server
$ git clone https://github.com/tensorflow/ecosystem.git
$ cd ecosystem/distribution_strategy
$ sudo nerdctl build --no-cache -t tf_std_server:v1 -f Dockerfile.tf_std_server . -namespace k8s.io
# check the built container image
$ sudo nerdctl image list --namespace k8s.io | grep tf_std_server
# export the built container image
$ sudo nerdctl save -o tf_std_server.tar.gz tf_std_server --namespace k8s.io
# scp the built container image to other nodes of the kubernetes cluster, where worker or parameter pod may be assigned.
$ scp tf_std_server.tar.gz maye@destinaton-node-ip:~
# import the built image on other nodes,
# in this example, the other node hasn't
# installed nerdctl, so use the built-in
# cli of containerd -- ctr.
"""
the default tag is docker.io/library/tf_std_server:v1
"""
$ sudo ctr -n k8s.io image import tf_std_server.tar.gz
Attention:
The . in sudo nerdctl build --no-cache -t tf_std_server:v1 -f Dockerfile.tf_std_server . -namespace k8s.io means specifying the current directory as the context, namely nerdctl build will find files it needs in this directory, if no directory specified, raise error: "FATA[0004] context needs to be specified " .
If not specifying namespace, the built image will be in namespace "default", and crictl (container runtime interface cli of kubernetes) can only see images in namespace "k8s.io" .
Note:
This directory contains the following files:
template.yaml.jinja: a jinja template to be rendered into a Kubernetes yaml file
Dockerfile.keras_model_to_estimator: a docker file to build the model image
Dockerfile.tf_std_server: a docker file to build the standard TensorFlow server image
keras_model_to_estimator.py: model code to run multi-worker training
tf_std_server.py: a standard TensorFlow binary
keras_model_to_estimator_client.py: model code to run in standalone client mode [1]
1.2 generate services worker, parameter server definition yaml file
1.2.1 Modify the header of jinja template: set image to tf_std_server:v1, script to /tf_std_server.py and cmdline_args to empty to run this standard TensorFlow server on each Kubernetes pod.
# file template.yaml.jinja
{%- set name = "dist-strat-example" -%}
{%- set image = "tf_std_server:v1" -%}
{%- set worker_replicas = 2 -%}
{%- set ps_replicas = 1 -%}
{%- set num_gpus_per_worker = 2 -%}
{%- set has_eval = False -%}
{%- set has_tensorboard = False -%}
{%- set train_dir = "gs://<your_gcs_bucket>" -%}
{%- set script = "/tf_std_server.py" -%}
{%- set cmdline_args = "" -%}
{%- set credential_secret_json = "key.json" -%}
{%- set credential_secret_key = "credential" -%}
{%- set port = 5000 -%}
{%- set replicas = {"worker": worker_replicas,
"ps": ps_replicas,
"evaluator": has_eval|int,
"tensorboard": has_tensorboard|int} -%}
{% set cmdline_arg_list = cmdline_args.split(" ") %}
{%- macro worker_hosts() -%}
{%- for i in range(worker_replicas) -%}
{%- if not loop.first -%},{%- endif -%}
\"{{ name }}-worker-{{ i }}:{{ port }}\"
{%- endfor -%}
{%- endmacro -%}
{%- macro ps_hosts() -%}
{%- for i in range(ps_replicas) -%}
{%- if not loop.first -%},{%- endif -%}
\"{{ name }}-ps-{{ i }}:{{ port }}\"
{%- endfor -%}
{%- endmacro -%}
{%- macro tf_config(task_type, task_id) -%}
{
\"cluster\": {
\"worker\": [{{ worker_hosts() }}]
{%- if ps_replicas > 0 -%}, \"ps\": [{{ ps_hosts() }}]{%- endif -%}
{%- if has_eval -%},
\"evaluator\": [\"{{ name }}-evaluator-0:{{ port }}\"]{%- endif -%}
},
\"task\": {
\"type\": \"{{ task_type }}\",
\"index\": \"{{ task_id }}\"
}
}
{%- endmacro -%}
{% for job in ["worker", "ps", "evaluator", "tensorboard"] -%}
{%- for i in range(replicas[job]) -%}
kind: Service
apiVersion: v1
metadata:
name: {{ name }}-{{ job }}-{{ i }}
spec:
type: LoadBalancer
selector:
name: {{ name }}
job: {{ job }}
task: "{{ i }}"
ports:
- port: {{ port }}
---
kind: ReplicationController
apiVersion: v1
metadata:
name: {{ name }}-{{ job }}-{{ i }}
spec:
replicas: 1
template:
metadata:
labels:
name: {{ name }}
job: {{ job }}
task: "{{ i }}"
spec:
containers:
{% if job == "tensorboard" %}
- name: tensorflow
image: tensorflow/tensorflow
{% else %}
- name: tensorflow
image: {{ image }}
resources:
limits:
nvidia.com/gpu: {{ num_gpus_per_worker }}
{% endif %}
env:
{% if job != "tensorboard" %}
- name: TF_CONFIG
value: "{{ tf_config(job, i) }}"
{% endif %}
- name: GOOGLE_APPLICATION_CREDENTIALS
value: "/var/secrets/google/{{ credential_secret_json }}"
ports:
- containerPort: {{ port }}
{% if job == "tensorboard" %}
command:
- "tensorboard"
args:
- "--logdir={{ train_dir }}"
- "--port={{ port }}"
{% else %}
command:
- "/usr/bin/python"
- "{{ script }}"
{%- for cmdline_arg in cmdline_arg_list %}
- "{{ cmdline_arg }}"
{%- endfor -%}
{% endif %}
volumeMounts:
- name: credential
mountPath: /var/secrets/google
volumes:
- name: credential
secret:
secretName: {{ credential_secret_key }}
---
{% endfor %}
{%- endfor -%}
1.2.2 modify tf_std_server.py, to start a tensorflow standard server (namely a worker or parameter server).
# file tf_std_server.py
# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
#
# 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.
# ==============================================================================
"""Run a standard tensorflow server."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
#import sys
import os
# This is for tensorflow v1, this example use
# tensorflow v2.
#def main(unused_argv):
# Contrib ops are lazily loaded. So we touch one contrib module to load them
# immediately.
#to_import_contrib_ops = tf.contrib.resampler
# Load you custom ops here before starting the standard TensorFlow server.
# Start and join the standard TensorFlow server.
#tf.contrib.distribute.run_standard_tensorflow_server().join()
# for tensorflow v2
def main():
cluster_resolver = tf.distribute.cluster_resolver.TFConfigClusterResolver()
if cluster_resolver.task_type in ("worker", "ps"):
# start s tensorflow server and wait.
os.environ["GRPC_FAIL_FAST"] = "use_caller"
server = tf.distribute.Server(
cluster_resolver.cluster_spec(),
job_name=cluster_resolver.task_type,
task_index=cluster_resolver.task_id,
protocol="grpc",
start=True)
server.join()
elif cluster_resolver.task_type == "evaluator":
# run sidecat evaluation
pass
else:
# run the coordinator
pass
if __name__ == "__main__":
#tf.app.run()
#main(sys.argv)
main()
1.2.3 generate services worker, parameter server definition yaml file
$ ../render_template.py template.yaml.jinja > worker_template.yaml
1.2.4 modify worker_template.yaml
# file worker_template.yaml
kind: Service
apiVersion: v1
metadata:
name: dist-strat-example-worker-0
# worker needs to mount pipeline_root directory,
# to access artifacts of tfx pipeline,
# pipeline_root directory is a persistentVolumeClaim
# in this example in namespace kubeflow, resource
# which wants to use the persistentVolumeClaim needs
# to be in the same namespace.
namespace: kubeflow
spec:
type: LoadBalancer
selector:
app: dist-strat-example-worker-0
ports:
- port: 5000
---
apiVersion: apps/v1
kind: Deployment
metadata:
# labels need to be at start of metadata,
# or will raise error.
labels:
app: dist-strat-example-worker-0
name: dist-strat-example-worker-0
namespace: kubeflow
spec:
replicas: 1
selector:
matchLabels:
app: dist-strat-example-worker-0
template:
metadata:
labels:
app: dist-strat-example-worker-0
spec:
# in this example, cpu of another node does not support
# AVX, which is needed by tensorflow v2, so
# use nodeAffinity to specify node this pod will be
# assigned to.
affinity:
nodeAffinity:
requiredDuringSchedulingIgnoredDuringExecution:
nodeSelectorTerms:
- matchExpressions:
- key: kubernetes.io/hostname
operator: In
values:
- maye-inspiron-5547
containers:
- name: tensorflow
image: tf_std_server:v1
resources:
limits:
# no gpu in this example, so comment out this line,
# set it according to the actual case.
#nvidia.com/gpu: 2
env:
# environmental varialbe TF_CONFIG, which contains
# information of the distributed training cluster.
# task is the role of this pod, tf_std_server.py
# will start tf.distribute.Server()
# based on TF_CONFIG.
- name: TF_CONFIG
value: "{
\"cluster\": {
\"worker\": [\"dist-strat-example-worker-0:5000\",\"dist-strat-example-worker-1:5000\"],
\"ps\": [\"dist-strat-example-ps-0:5000\"]},
\"task\": {
\"type\": \"worker\",
\"index\": \"0\"
}
}"
#- name: GOOGLE_APPLICATION_CREDENTIALS
# value: "/var/secrets/google/key.json"
ports:
- containerPort: 5000
command:
- "/usr/bin/python"
- "/tf_std_server.py"
- ""
volumeMounts:
- mountPath: /tfx/tfx_pv
name: tfx-pv
#- name: credential
# mountPath: /var/secrets/google
volumes:
- name: tfx-pv
persistentVolumeClaim:
claimName: tfx-pv-claim
#- name: credential
# secret:
# secretName: credential
---
kind: Service
apiVersion: v1
metadata:
name: dist-strat-example-worker-1
namespace: kubeflow
spec:
type: LoadBalancer
selector:
app: dist-strat-example-worker-1
ports:
- port: 5000
---
apiVersion: apps/v1
kind: Deployment
metadata:
labels:
app: dist-strat-example-worker-1
name: dist-strat-example-worker-1
namespace: kubeflow
spec:
replicas: 1
selector:
matchLabels:
app: dist-strat-example-worker-1
template:
metadata:
labels:
app: dist-strat-example-worker-1
spec:
affinity:
nodeAffinity:
requiredDuringSchedulingIgnoredDuringExecution:
nodeSelectorTerms:
- matchExpressions:
- key: kubernetes.io/hostname
operator: In
values:
- maye-inspiron-5547
containers:
- name: tensorflow
image: tf_std_server:v1
resources:
limits:
#nvidia.com/gpu: 2
env:
- name: TF_CONFIG
value: "{
\"cluster\": {
\"worker\": [\"dist-strat-example-worker-0:5000\",\"dist-strat-example-worker-1:5000\"],
\"ps\": [\"dist-strat-example-ps-0:5000\"]},
\"task\": {
\"type\": \"worker\",
\"index\": \"1\"
}
}"
#- name: GOOGLE_APPLICATION_CREDENTIALS
# value: "/var/secrets/google/key.json"
ports:
- containerPort: 5000
command:
- "/usr/bin/python"
- "/tf_std_server.py"
- ""
volumeMounts:
- mountPath: /tfx/tfx_pv
name: tfx-pv
#volumeMounts:
#- name: credential
# mountPath: /var/secrets/google
volumes:
- name: tfx-pv
persistentVolumeClaim:
claimName: tfx-pv-claim
#volumes:
#- name: credential
# secret:
# secretName: credential
---
kind: Service
apiVersion: v1
metadata:
name: dist-strat-example-ps-0
namespace: kubeflow
spec:
type: LoadBalancer
selector:
app: dist-strat-example-ps-0
ports:
- port: 5000
---
apiVersion: apps/v1
kind: Deployment
metadata:
labels:
app: dist-strat-example-ps-0
name: dist-strat-example-ps-0
namespace: kubeflow
spec:
replicas: 1
selector:
matchLabels:
app: dist-strat-example-ps-0
template:
metadata:
labels:
app: dist-strat-example-ps-0
spec:
affinity:
nodeAffinity:
requiredDuringSchedulingIgnoredDuringExecution:
nodeSelectorTerms:
- matchExpressions:
- key: kubernetes.io/hostname
operator: In
values:
- maye-inspiron-5547
containers:
- name: tensorflow
image: tf_std_server:v1
resources:
limits:
#nvidia.com/gpu: 2
env:
- name: TF_CONFIG
value: "{
\"cluster\": {
\"worker\": [\"dist-strat-example-worker-0:5000\",\"dist-strat-example-worker-1:5000\"],
\"ps\": [\"dist-strat-example-ps-0:5000\"]},
\"task\": {
\"type\": \"ps\",
\"index\": \"0\"
}
}"
#- name: GOOGLE_APPLICATION_CREDENTIALS
# value: "/var/secrets/google/key.json"
ports:
- containerPort: 5000
command:
- "/usr/bin/python"
- "/tf_std_server.py"
- ""
volumeMounts:
- mountPath: /tfx/tfx_pv
name: tfx-pv
#volumeMounts:
#- name: credential
# mountPath: /var/secrets/google
volumes:
- name: tfx-pv
persistentVolumeClaim:
claimName: tfx-pv-claim
#volumes:
#- name: credential
# secret:
# secretName: credential
---
1.3 deploy services worker, parameter server
$ kubectl apply -f worker_template.yaml
2. use tf.distribute.ParameterServerStrategy() in run_fn() in module file of tfx component trainer.
def run_fn(fn_args: tfx.components.FnArgs):
cluster_dict = {}
#cluster_dict["worker"] = ["dist-strat-example-worker-0:5000", "dist-strat-example-worker-1:5000"]
#cluster_dict["ps"] = ["dist-strat-example-ps-0:5000"]
### need to use ClusterIP of services worker, ps,
### not service name, or can not connect to them.
cluster_dict["worker"] = ["10.105.206.29:5000", "10.102.137.138:5000"]
cluster_dict["ps"] = ["10.105.27.97:5000"]
cluster_spec = tf.train.ClusterSpec(cluster_dict)
cluster_resolver = tf.distribute.cluster_resolver.SimpleClusterResolver(
cluster_spec, rpc_layer="grpc")
### distribution strategy needs to be instantiated at start
### of run_fn(), before any tensorflow operation.
strategy = tf.distribute.ParameterServerStrategy(
cluster_resolver,)
tf_transform_output = tft.TFTransformOutput(fn_args.transform_output)
#tf.print(f"fn_args.train_files: {fn_args.train_files}")
train_dataset = _input_fn(
fn_args.train_files,
fn_args.data_accessor,
tf_transform_output,
batch_size=_TRAIN_BATCH_SIZE,
)
resampled_train_dataset = _resample_train_dataset(train_dataset,
batch_size=_TRAIN_BATCH_SIZE)
#tf.print(f"resampled_train_dataset {resampled_train_dataset.cardinality()}")
val_dataset = _input_fn(
fn_args.eval_files,
fn_args.data_accessor,
tf_transform_output,
batch_size=_EVAL_BATCH_SIZE,
)
### val_dataset needs to be infinite, or will raise error:
### OutOfRange.
val_dataset = val_dataset.repeat()
#tf.print(f"val_dataset cardinality: {val_dataset.cardinality()}")
### instantiate model, metric, optimizer,loss, compile model
### in the scope of distribution strategy.
with strategy.scope():
model = _build_keras_model()
#tf.print(f"custom_config: {fn_args.custom_config}")
#log_dir = "logs/fit/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
#tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir,)
### callbacks.BackupAndRestore() is for the case when
### training cluster is not available temporally,
### continue the training from the interrupted epoch.
backup_dir = os.path.join("/home/maye/maye_temp", "backup")
callbacks = [
tf.keras.callbacks.BackupAndRestore(backup_dir=backup_dir),
]
trainer_train_history = model.fit(
resampled_train_dataset,
epochs=fn_args.custom_config['epochs'],
steps_per_epoch=fn_args.train_steps,
validation_data=val_dataset,
### since val_dataset is infinite, validation_steps needs
### to be specified.
validation_steps=3,
callbacks=callbacks,
)
#tf.print(f"train_history: \n {train_history.history}")
with open('trainer_train_history.json', 'w') as f:
json.dump(trainer_train_history.history, f)
#signatures = {
# 'serving_default': _get_serve_tf_examples_fn(model, tf_transform_output),
#}
#model.save(fn_args.serving_model_dir, save_format='tf', signatures=signatures)
Check ClusterIp of service:
(base) maye@maye-Inspiron-5547:~/github_repository/tensorflow_ecosystem/distribution_strategy$ kubectl get service -n kubeflow
NAME TYPE CLUSTER-IP EXTERNAL-IP PORT(S) AGE
dist-strat-example-ps-0 LoadBalancer 10.96.200.160 <pending> 5000:32409/TCP 53m
dist-strat-example-worker-0 LoadBalancer 10.102.74.8 <pending> 5000:30550/TCP 53m
dist-strat-example-worker-1 LoadBalancer 10.100.198.218 <pending> 5000:31080/TCP 53m
3. run the tfx pipeline using kubeflow pipeline, refer to «Run a tfx pipeline using kubeflow pipeline» https://www.cnblogs.com/zhenxia-jiuyou/p/18003167
ok log of pod tfx-component-trainer:
Epoch 1/50
/usr/local/lib/python3.8/dist-packages/tensorflow/python/data/ops/dataset_ops.py:467: UserWarning: To make it possible to preserve tf.data options across serialization boundaries, their implementation has moved to be part of the TensorFlow graph. As a consequence, the options value is in general no longer known at graph construction time. Invoking this method in graph mode retains the legacy behavior of the original implementation, but note that the returned value might not reflect the actual value of the options.
warnings.warn("To make it possible to preserve tf.data options across "
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Reduce to /device:CPU:0 then broadcast to ('/replica:0/device:CPU:0',).
INFO:tensorflow:Waiting for all global closures to be finished.
INFO:tensorflow:Waiting for all global closures to be finished.
INFO:tensorflow:Waiting for all global closures to be finished.
INFO:tensorflow:Waiting for all global closures to be finished.
21/21 - 31s - loss: 0.7040 - cross entropy: 0.7036 - tp: 882.0000 - fp: 820.0000 - tn: 547.0000 - fn: 439.0000 - precision: 0.5182 - recall: 0.6677 - auc: 0.5415 - prc: 0.5361 - val_loss: 0.6753 - val_cross entropy: 0.6749 - val_tp: 30.0000 - val_fp: 181.0000 - val_tn: 166.0000 - val_fn: 7.0000 - val_precision: 0.1422 - val_recall: 0.8108 - val_auc: 0.7620 - val_prc: 0.3268 - 31s/epoch - 1s/step
Epoch 2/50
INFO:tensorflow:Waiting for all global closures to be finished.
INFO:tensorflow:Waiting for all global closures to be finished.
INFO:tensorflow:Waiting for all global closures to be finished.
INFO:tensorflow:Waiting for all global closures to be finished.
21/21 - 14s - loss: 0.5821 - cross entropy: 0.5817 - tp: 1036.0000 - fp: 439.0000 - tn: 926.0000 - fn: 287.0000 - precision: 0.7024 - recall: 0.7831 - auc: 0.8096 - prc: 0.8139 - val_loss: 0.5677 - val_cross entropy: 0.5673 - val_tp: 25.0000 - val_fp: 82.0000 - val_tn: 271.0000 - val_fn: 6.0000 - val_precision: 0.2336 - val_recall: 0.8065 - val_auc: 0.8646 - val_prc: 0.3799 - 14s/epoch - 657ms/step
Epoch 3/50
INFO:tensorflow:Waiting for all global closures to be finished.
INFO:tensorflow:Waiting for all global closures to be finished.
INFO:tensorflow:Waiting for all global closures to be finished.
INFO:tensorflow:Waiting for all global closures to be finished.
Note:
1.
Keras Model.fit with parameter server training assumes that each worker receives the same dataset, except when it is shuffled differently. Therefore, by calling Dataset.shuffle, you ensure more even iterations over the data.
Because workers do not synchronize, they may finish processing their datasets at different times. Therefore, the easiest way to define epochs with parameter server training is to use Dataset.repeat—which repeats a dataset indefinitely when called without an argument—and specify the steps_per_epoch argument in the Model.fit call.
2. Parameter server training is a common data-parallel method to scale up model training on multiple machines. A parameter server training cluster consists of workers and parameter servers.
Variables are created on parameter servers and they are read and updated by workers in each step.
3. the worker and parameter server tasks run tf.distribute.Servers that listen for tasks from the coordinator. The coordinator creates resources, dispatches training tasks, writes checkpoints, and deals with task failures.
4. You will need to configure the 'TF_CONFIG' environment variable if you use TFConfigClusterResolver.
5. Even if you choose the Model.fit training path, you can optionally instantiate a tf.distribute.coordinator.ClusterCoordinator object to schedule other functions you would like to be executed on the workers.
6. In TensorFlow 2, parameter server training is powered by the tf.distribute.ParameterServerStrategy class, which distributes the training steps to a cluster that scales up to thousands of workers (accompanied by parameter servers).
7. When using parameter server training, it is recommended to have:
One coordinator job (which has the job name chief)
Multiple worker jobs (job name worker)
Multiple parameter server jobs (job name ps)
In this example, the pod trainer is the coordinator job, the process where tf.distribute.ParameterServerStrategy() is called is the coordinator job, no need to specify the coordinator job in the cluster_dict specially.
- The coordinator creates resources, dispatches training tasks, writes checkpoints, and deals with task failures. The workers and parameter servers run tf.distribute.Server instances that listen for requests from the coordinator.
- a parameter server training cluster requires a coordinator task that runs your training program, one or several workers and parameter server tasks that run TensorFlow servers—tf.distribute.Server.
The requirements to set them up are:
The coordinator task needs to know the addresses and ports of all other TensorFlow servers, except the evaluator.
The workers and parameter servers need to know which port they need to listen to. For the sake of simplicity, you can usually pass in the complete cluster information when creating TensorFlow servers on these tasks.
Workers and parameter servers should have task types as "worker" and "ps", respectively. The coordinator should use "chief" as the task type for legacy reasons.
You will start by creating several TensorFlow servers in advance and you will connect to them later.
[2]
When ClusterIPs or hostnames of workers, parameter servers are passed to tf.distribute.ParameterServerStrategy(cluster_resolver) of coordinator correctly, and workers, parameter servers are ready, the coordinator will connect to workers, parameter servers and start distributed training, the training log will be shown in log of the coordinator.
4. Error & Solution
[ERRRO: AttributeError: module 'tensorflow' has no attribute 'app']
(base) maye@maye-Inspiron-5547:~/github_repository/tensorflow_ecosystem/distribution_strategy$ kubectl logs dist-strat-example-worker-0-w6rsb
2024-02-03 07:38:33.104872: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
Traceback (most recent call last):
File "/tf_std_server.py", line 35, in
tf.app.run()
^^^^^^
AttributeError: module 'tensorflow' has no attribute 'app'
[SOLUTION]
This is due to that the tensorflow in use is v2, tf.app.run() is a sentence of tensorflow v1, module app has been removed in tensorflow v2,
tf.app.run() = main(sys.argv)
Use tf.distribute.Server() to start a tensorflow server for tensorflow v2.
Note:
"exit code: 1" : something wrong in executing code of the process.
"exit code: 137": the process has received SIGNAL KILL, in the case of kubernetes, if kubelet needs to stop a container process, it will call containerd, and containerd will send SIGNAL KILL to the container process. Linux will send SIGNAL KILL to a process if cpu, or memory is not enough.
"exit code: 139': SEGMENT FAULT, the process tries to access memory, or file, or table in a database which is not accessible, such as, memory out of boundary, not existed file or database table.
References: