[HDFS-9259] Make SO_SNDBUF size configurable at DFSClient side for hdfs write scenario

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 2.8.0, 3.0.0-alpha1
• Hadoop Flags: Reviewed
• Release Note: Introduces a new configuration setting dfs.client.socket.send.buffer.size to control the socket send buffer size for writes. Setting it to zero enables TCP auto-tuning on systems that support it.
• Assignee: Mingliang Liu
• Reporter: Ming Ma
• Created: 16/Oct/15 15:49
• Updated: 16/Nov/17 04:35
• Resolved: 27/Oct/15 16:33

Description

We recently found that cross-DC hdfs write could be really slow. Further investigation identified that is due to SendBufferSize and ReceiveBufferSize used for hdfs write. The test ran "hadoop -fs -copyFromLocal" of a 256MB file across DC with different SendBufferSize and ReceiveBufferSize values. The results showed that c much faster than b; b is faster than a.

a. SendBufferSize=128k, ReceiveBufferSize=128k (hdfs default setting).
b. SendBufferSize=128K, ReceiveBufferSize=not set(TCP auto tuning).
c. SendBufferSize=not set, ReceiveBufferSize=not set(TCP auto tuning for both)

HDFS-8829 has enabled scenario b. We would like to enable scenario c by making SendBufferSize configurable at DFSClient side. Cc: cmccabe He Tianyi kanaka vinayrpet.

Issue Links

• is related to: HDFS-10326 Disable setting tcp socket send/receive buffers for write pipelines
• relates to: HDFS-8829 Make SO_RCVBUF and SO_SNDBUF size configurable for DataTransferProtocol sockets and allow configuring auto-tuning
• relates to: HDFS-12823 Backport HDFS-9259 "Make SO_SNDBUF size configurable at DFSClient" to branch-2.7