[ https://issues.apache.org/jira/browse/GIRAPH-1056?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15258565#comment-15258565 ]
Hassan Eslami commented on GIRAPH-1056: --------------------------------------- Thanks Maja for reviewing and committing the diff. > Setting auto-read in Netty to false > ----------------------------------- > > Key: GIRAPH-1056 > URL: https://issues.apache.org/jira/browse/GIRAPH-1056 > Project: Giraph > Issue Type: Improvement > Reporter: Hassan Eslami > Assignee: Hassan Eslami > Priority: Minor > > By default, auto-read flag is set to true in Netty. This means Netty > proactively read requests as they become available to a worker. However, this > behavior sometime causes the off-heap memory to increase continuously. This > happens specifically in presence of a spike in the amount of received > requests. In that situation, the processing/handling rate of incoming > requests may be less than the request receipt rate leading to high-memory > kill (CGroup kill or OOM). With auto-read flag set to false, we read and > process requests one by one and (hopefully/presumably) letting the transport > layer do the flow control (i.e. dropping packets or reducing congestion > window of TCP). -- This message was sent by Atlassian JIRA (v6.3.4#6332)