[
https://issues.apache.org/jira/browse/HADOOP-6577?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12835552#action_12835552
]
Suresh Srinivas commented on HADOOP-6577:
-----------------------------------------
When a high frequency of requests are made to namenode, each resulting in a
response of size greater than 1MB, a lot of garbage is created on the heap.
This could result in tenured heap getting filled up very fast and trigger full
GC. Full GC results in long stop the world pauses, affecting the applications
using HDFS.
In one of the instances observed on the production cluster, an application
repeatedly made list status calls, each with response sizes ranging from 3 to
5MB. This resulted in full GC, which could have been avoided by setting the max
response size to 10MB.
More permanent solution for this problem is to ensure an operation that results
in large response (listStatus) are broken in to multi-step smaller operations.
This will be addressed in a separate jira. In the interim I propose adding a
hidden config param that could be used for setting the max resp buffer size.
> IPC server response buffer reset threshold should be configurable
> -----------------------------------------------------------------
>
> Key: HADOOP-6577
> URL: https://issues.apache.org/jira/browse/HADOOP-6577
> Project: Hadoop Common
> Issue Type: Bug
> Affects Versions: 0.22.0
> Reporter: Suresh Srinivas
> Assignee: Suresh Srinivas
> Fix For: 0.22.0
>
>
> In HDFS-6460, the response buffers in o.a.h.ipc.Server.Handler was reset when
> the buffer grows beyond max size of 1MB. This frees heap from large responses
> occupying it. This max response size limit should be configurable. Details in
> subsequent comment.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
