[YARN-2681] Support bandwidth enforcement for containers while reading from HDFS

• Type: New Feature
• Status: Patch Available
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.5.1
• Component/s: nodemanager

Environment

Linux

• Target Version/s: 3.5.0
• Release Note: support HDFS read bandwidth enforcement in case data is tranfered over TCP/IP socket
• Assignee: Unassigned
• Reporter: Nam H. Do
• Created: 13/Oct/14 11:05
• Updated: 04/Jan/24 09:24

Description

To read/write data from HDFS on data node, applications establise TCP/IP connections with the datanode. The HDFS read can be controled by setting Linux Traffic Control (TC) subsystem on the data node to make filters on appropriate connections.

The current cgroups net_cls concept can not be applied on the node where the container is launched, netheir on data node since:

• TC hanldes outgoing bandwidth only, so it can be set on container node (HDFS read = incoming data for the container)
• Since HDFS data node is handled by only one process, it is not possible to use net_cls to separate connections from different containers to the datanode.

Tasks:
1) Extend Resource model to define bandwidth enforcement rate
2) Monitor TCP/IP connection estabilised by container handling process and its child processes
3) Set Linux Traffic Control rules on data node base on address:port pairs in order to enforce bandwidth of outgoing data

Concept: http://www.hit.bme.hu/~do/papers/EnforcementDesign.pdf
Implementation:
http://www.hit.bme.hu/~dohoai/documents/HdfsTrafficControl.pdf
http://www.hit.bme.hu/~dohoai/documents/HdfsTrafficControl_UML_diagram.png

Issue Links

• is related to: YARN-3336 FileSystem memory leak in DelegationTokenRenewer
• relates to: YARN-3926 [Umbrella] Extend the YARN resource model for easier resource-type management and profiles