当前性能问题描述

1. 增加worker数，一定范围内能带来较好的性能提升，但是继续增加worker数时，训练性能提升不明显；
2. 增加ps数，一定范围内能带来较好的性能提升，但是继续增加ps数时，训练性能提升不明显；

可能原因：

1. 与ps和worker的分布情况强相关：

   • 目前的调度策略，主要根据服务器的cpu和内存使用情况进行均衡调度，尽量使得集群中每台服务器的cpu和内存使用率相当。这种情况下，ps和worker的调度存在一定程度的随机性。
   • 如果调度时，每台包含worker的服务器都有对应一个ps，那么训练性能会更高?如果有，性能提升多少呢？

2. K8S中的worker从HDFS集群中读取训练数据时存在IO瓶颈？可能网络上的或者是HDFS本身的配置，需要通过HDFS集群的监控来进一步排查。

下面，是针对第一种“可能原因：与ps和worker的分布情况强相关“ 设计的测试场景和用例：

场景1：将每个worker所在的服务器都有对应的ps。

测试用例

TensorFlow tasks调度设计图

调度实现

• 场景1的TensorFlow对象模板***scene1.jinja***

# scene1.jinja —— 对象模板
{%- set name = "##NAME##" -%}
{%- set worker_replicas = ##WN## -%}
{%- set ps_replicas = ##PN## -%}
{%- set script = "##SCRIPT##" -%}
{%- set case = "##CASE##" -%}


{%- set port = 2222 -%}
{%- set log_host_dir = "/var/log/tensorflow" -%}
{%- set log_container_dir = "/var/log" -%}
{%- set image = "registry.vivo.xyz:4443/bigdata_release/tensorflow1.3.0" -%}
{%- set replicas = {"worker": worker_replicas, "ps": ps_replicas} -%}

{%- 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 -%}


{%- for i in range( begin_index, end_index ) -%}
{%- if task_type == "worker" %}

---
kind: Service
apiVersion: v1
metadata:
 name: {{ name }}-{{ task_type }}-{{ i }}
 namespace: {{ name }}
spec:
 clusterIP: None
 selector:
 name: {{ name }}
 job: {{ task_type }}
 task: "{{ i }}"
 ports:
 - port: {{ port }}
 targetPort: 2222
---
kind: Job
apiVersion: batch/v1
metadata:
 name: {{ name }}-{{ task_type }}-{{ i }}
 namespace: {{ name }}
spec:
 template:
 metadata:
 labels:
 name: {{ name }}
 job: {{ task_type }}
 task: "{{ i }}"
 spec:
 imagePullSecrets:
 - name: harborsecret'
 affinity:
 nodeAffinity:
 requiredDuringSchedulingIgnoredDuringExecution:
 nodeSelectorTerms:
 - matchExpressions:
 - key: "CASE" operator: In
 values: 
 - "{{ case }}"
 - key: "INDEX" operator: In
 values: 
 - "{{ i // 10 }}"
 - key: "SCENCE" operator: In
 values: 
 - "1"
 containers:
 - name: {{ name }}-{{ task_type }}-{{ i }}
 image: {{ image }}
 resources:
 requests:
 memory: "4Gi"
 cpu: "300m"
 ports:
 - containerPort: 2222
 command: ["/bin/sh", "-c", "export CLASSPATH=.:/usr/lib/jvm/java-1.8.0/lib/tools.jar:$(/usr/lib/hadoop-2.6.1/bin/hadoop classpath --glob); wget -r -nH -np --cut-dir=1 -R 'index.html*,*gif' {{ script }}; cd ./{{ name }}; sh ./run.sh {{ ps_hosts() }} {{ worker_hosts() }} {{ task_type }} {{ i }} {{ ps_replicas }} {{ worker_replicas }}"]
 restartPolicy: OnFailure

{%- endif -%}

{%- if task_type == "ps" -%}
---
kind: Service
apiVersion: v1
metadata:
 name: {{ name }}-{{ task_type }}-{{ i }}
 namespace: {{ name }}
spec:
 clusterIP: None
 selector:
 name: {{ name }}
 job: {{ task_type }}
 task: "{{ i }}"
 ports:
 - port: {{ port }}
 targetPort: 2222
---
kind: Deployment
apiVersion: extensions/v1beta1
metadata:
 name: {{ name }}-{{ task_type }}-{{ i }}
 namespace: {{ name }}
spec:
 replicas: 1 template:
 metadata:
 labels:
 name: {{ name }}
 job: {{ task_type }}
 task: "{{ i }}"
 spec:
 imagePullSecrets:
 - name: harborsecret
 affinity:
 nodeAffinity:
 requiredDuringSchedulingIgnoredDuringExecution:
 nodeSelectorTerms:
 - matchExpressions:
 - key: "CASE" operator: In
 values: 
 - "{{ case }}"
 - key: "INDEX" operator: In
 values: 
 - "{{ i }}"
 - key: "SCENCE" operator: In
 values: 
 - "1"
 containers:
 - name: {{ name }}-{{ task_type }}-{{ i }}
 image: {{ image }}
 resources:
 requests:
 memory: "4Gi"
 cpu: "2"
 ports:
 - containerPort: 2222
 command: ["/bin/sh", "-c","export CLASSPATH=.:/usr/lib/jvm/java-1.8.0/lib/tools.jar:$(/usr/lib/hadoop-2.6.1/bin/hadoop classpath --glob); wget -r -nH -np --cut-dir=1 -R 'index.html*,*gif' {{ script }}; cd ./{{ name }}; sh ./run.sh {{ ps_hosts() }} {{ worker_hosts() }} {{ task_type }} {{ i }} {{ ps_replicas }} {{ worker_replicas }}"]
 restartPolicy: Always
{%- endif -%}
{%- endfor -%}

• Label Nodes

选择对应的节点打上对应的Label。

kubectl label node $node_name SCENCE=1 CASE=? INDEX=?

测试结果

用例2的测试截图：

场景2：将所有ps和所有worker都强制进行物理隔离。

测试用例

TensorFlow tasks调度设计图

调度实现

• 场景2的TensorFlow对象模板***scene2.jinja***

# scene2.jinja —— 对象模板
{%- set name = "##NAME##" -%}
{%- set worker_replicas = ##WN## -%}
{%- set ps_replicas = ##PN## -%}
{%- set script = "##SCRIPT##" -%}
{%- set case = "##CASE##" -%}


{%- set port = 2222 -%}
{%- set log_host_dir = "/var/log/tensorflow" -%}
{%- set log_container_dir = "/var/log" -%}
{%- set image = "registry.vivo.xyz:4443/bigdata_release/tensorflow1.3.0" -%}
{%- set replicas = {"worker": worker_replicas, "ps": ps_replicas} -%}

{%- 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 -%}


{%- for i in range( begin_index, end_index ) -%}
{%- if task_type == "worker" %}

---
kind: Service
apiVersion: v1
metadata:
 name: {{ name }}-{{ task_type }}-{{ i }}
 namespace: {{ name }}
spec:
 clusterIP: None
 selector:
 name: {{ name }}
 job: {{ task_type }}
 task: "{{ i }}"
 ports:
 - port: {{ port }}
 targetPort: 2222
---
kind: Job
apiVersion: batch/v1
metadata:
 name: {{ name }}-{{ task_type }}-{{ i }}
 namespace: {{ name }}
spec:
 template:
 metadata:
 labels:
 name: {{ name }}
 job: {{ task_type }}
 task: "{{ i }}"
 spec:
 imagePullSecrets:
 - name: harborsecret'
 affinity:
 nodeAffinity:
 requiredDuringSchedulingIgnoredDuringExecution:
 nodeSelectorTerms:
 - matchExpressions:
 - key: "CASE" operator: In
 values: 
 - "{{ case }}"
 - key: "INDEX" operator: In
 values: 
 - "{{ i // 10 }}"
 - key: "SCENCE" operator: In
 values: 
 - "2"
 - key: "TYPE" operator: In
 values: 
 - "worker"
 containers:
 - name: {{ name }}-{{ task_type }}-{{ i }}
 image: {{ image }}
 resources:
 requests:
 memory: "4Gi"
 cpu: "300m"
 ports:
 - containerPort: 2222
 command: ["/bin/sh", "-c", "export CLASSPATH=.:/usr/lib/jvm/java-1.8.0/lib/tools.jar:$(/usr/lib/hadoop-2.6.1/bin/hadoop classpath --glob); wget -r -nH -np --cut-dir=1 -R 'index.html*,*gif' {{ script }}; cd ./{{ name }}; sh ./run.sh {{ ps_hosts() }} {{ worker_hosts() }} {{ task_type }} {{ i }} {{ ps_replicas }} {{ worker_replicas }}"]
 restartPolicy: OnFailure

{%- endif -%}

{%- if task_type == "ps" -%}
---
kind: Service
apiVersion: v1
metadata:
 name: {{ name }}-{{ task_type }}-{{ i }}
 namespace: {{ name }}
spec:
 clusterIP: None
 selector:
 name: {{ name }}
 job: {{ task_type }}
 task: "{{ i }}"
 ports:
 - port: {{ port }}
 targetPort: 2222
---
kind: Deployment
apiVersion: extensions/v1beta1
metadata:
 name: {{ name }}-{{ task_type }}-{{ i }}
 namespace: {{ name }}
spec:
 replicas: 1 template:
 metadata:
 labels:
 name: {{ name }}
 job: {{ task_type }}
 task: "{{ i }}"
 spec:
 imagePullSecrets:
 - name: harborsecret
 affinity:
 nodeAffinity:
 requiredDuringSchedulingIgnoredDuringExecution:
 nodeSelectorTerms:
 - matchExpressions:
 - key: "CASE" operator: In
 values: 
 - "{{ case }}"
 - key: "INDEX" operator: In
 values: 
 - "{{ i }}"
 - key: "SCENCE" operator: In
 values: 
 - "2"
 - key: "TYPE" operator: In
 values: 
 - "ps"
 containers:
 - name: {{ name }}-{{ task_type }}-{{ i }}
 image: {{ image }}
 resources:
 requests:
 memory: "4Gi"
 cpu: "2"
 ports:
 - containerPort: 2222
 command: ["/bin/sh", "-c","export CLASSPATH=.:/usr/lib/jvm/java-1.8.0/lib/tools.jar:$(/usr/lib/hadoop-2.6.1/bin/hadoop classpath --glob); wget -r -nH -np --cut-dir=1 -R 'index.html*,*gif' {{ script }}; cd ./{{ name }}; sh ./run.sh {{ ps_hosts() }} {{ worker_hosts() }} {{ task_type }} {{ i }} {{ ps_replicas }} {{ worker_replicas }}"]
 restartPolicy: Always
{%- endif -%}
{%- endfor -%}

• Label Nodes

选择对应的节点打上对应的Label。

kubectl label node $node_name SCENCE=1 CASE=? INDEX=? TYPE=?

测试结果

用例2的测试截图：

测试结论及思考

对比两种不同场景下用例2（5个ps，50个worker）的监控数据，发现如下现象：

• 两种场景下，虽然创建了5个ps，但是实际上只有一个ps的负载比较高，其他的ps要么cpu usage在10%以下，要么甚至几乎为0。

• 两种场景下同样的worker number和ps number，整个tensorflow cluster消耗的cpu和内存差别很小。

测试结论

• 分布式tensorflow中，每个worker选择哪个ps作为自己的参数服务器跟我们如何强制分布ps和worker的布局无关，由分布式tensorflow内部自己控制（跟tf.train.replica_device_setter()设置的strategy有关）。

问题思考

• 为什么这个训练中，多个ps中只有一个ps在工作？是算法只有一个Big参数？如果是，那么默认按照Round-Robin策略只会使用一个ps，就能解释这个问题了。这需要算法的兄弟进行确认。

• 如果将Big参数拆分成众多Small参数，使用RR或LB或Partition策略之一，应该都能利用多个ps进行参数更新明显提升训练性能。

• 通过这次折腾，也不是一无所获，至少发现我们对于Distributed TensorFlow的内部工作原理还不甚了解，非常有必要深入到源码进行分析。