提交 68c19943 编写于 作者: M m3ngyang

fix typo

上级 efd7ee85
...@@ -4,13 +4,13 @@ Cluster Training and Prediction ...@@ -4,13 +4,13 @@ Cluster Training and Prediction
.. contents:: .. contents::
1. Network connection errors in the log during muliti-node cluster training 1. Network connection errors in the log during multi-node cluster training
------------------------------------------------ ------------------------------------------------
The errors in the log belong to network connection during mulilti-node cluster training, for example, :code:`Connection reset by peer`. There are maybe some errors in the log belonging to network connection problem during multi-node cluster training, for example, :code:`Connection reset by peer`.
This kind of error is usually caused by the abnormal exit of the training process in some node, and the others cannot connect with this node any longer. Steps to troubleshoot the problem as follows: This kind of error is usually caused by the abnormal exit of a training process in some node, and the other nodes cannot connect with this node any longer. Steps to troubleshoot the problem are as follows:
* Find the first error in the :code:`train.log`, :code:`server.log`, check whether other fault casued the problem, such as FPE, lacking of memory or disk. * Find the first error in the :code:`train.log`, :code:`server.log`, check whether other fault casued the problem, such as FPE, lacking of memory or disk.
* If network connection gave rise to the first error in the log, this may be caused by the port conflict of the non-exclusive execution. Connect with the operator to check if the current MPI cluster supports jobs submitted with parameter :code:`resource=full`. If so, change the port of job. * If the first error in server.log says "Address already used", this may be caused by the port conflict of the non-exclusive execution. Connect the sys-admin to check if the current MPI cluster supports jobs submitted with parameter :code:`resource=full`. If the current MPI cluster does not support this parameter, change the server port and try agian.
* If the currnet MPI cluster does not support exclusive pattern, ask the operator to replace or update the current cluster. * If the current MPI cluster does not support exclusive pattern which allows a process to occupy the whole node, ask the administrator to replace or update the this cluster.
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册