Tensorboard跳板机端口转发配置

出于网络安全策略考虑，内网服务器无法连接外网，一般通过跳板机免密登录，因此tensorboard无法使用。本文介绍通过SSH隧道技术穿透内外网限制，实现tensorboard实时监测tensorflow深度学习模型运行。

 

1、目标服务器host-inaccessible建立root权限的新用户（知晓root密码的可以跳过此步）

#建立新用户yzgpu，并添加到root用户的group(root)
[root@100-020-gpuserver ~]# adduser -g root yzgpu
#设置yzgpu密码
[root@100-020-gpuserver ~]# passwd yzgpu

#彻底删除某个用户，包括/home/yzgpu和邮件池/var/spool/mail/yzgpu，防止建立相同的用户报错
#[root@100-020-gpuserver ~]# userdel -r yzgpu
#

2、拥有等同root权限（root）的能与外网连接的服务器host-accessible，或者跳板机服务器host-jump

3、建立host-inaccessible和host-accessible（host-jump）之间的SSH隧道（方法见本人博客：SSH隧道技术-实现内网不可访问的服务器与外网互访）

#host-accessible服务器运行
[root@hdfs08-dev yzadmin]# ssh -C -f -N -g -L 1817:host-inaccessible-ip:6006 host-inaccessible-username@host-inaccessible-ip

注意：将其中的端口号改为tensorboard指定端口号，若未指定则为6006即可

4、host-inaccessible服务器运行tensorboard --logdir=eval

[yzgpu@100-020-gpuserver faster_rcnn]$ tensorboard --logdir=eval
W0917 14:09:46.598918 139717303437056 plugin_event_accumulator.py:294] Found more than one graph event per run, or there was a metagraph containing a graph_def, as well as one or more graph events.  Overwriting the graph with the newest event.
W0917 14:09:46.603488 139717303437056 plugin_event_accumulator.py:302] Found more than one metagraph event per run. Overwriting the metagraph with the newest event.
W0917 14:09:46.653152 139717303437056 plugin_event_accumulator.py:294] Found more than one graph event per run, or there was a metagraph containing a graph_def, as well as one or more graph events.  Overwriting the graph with the newest event.
W0917 14:09:46.659667 139717303437056 plugin_event_accumulator.py:302] Found more than one metagraph event per run. Overwriting the metagraph with the newest event.
TensorBoard 1.14.0 at http://host-inaccessible-ip:6006/ (Press CTRL+C to quit)

5、打开host-accessible服务器，网页输入http://host-accessible-ip:1817，即可运行tensorboard可视化界面

(若该网址无法访问，可使用http://host-inaccessible-ip:6006，两者网址实际指向相同，但本人实际操作发现初次使用http://host-accessible-ip:1817有效，后面后续使用http://host-inaccessible-ip:6006才是有效连接，原因不明)