[HDFS-12077] Implement a remaining space based balancer policy

• Type: New Feature
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.6.0
• Component/s: balancer & mover
• Assignee: Unassigned
• Reporter: liuyiyang
• Created: 02/Jul/17 14:15
• Updated: 05/Oct/17 20:35

Description

Our cluster has DataNodes with 2T disk storage, as storage utilization of the cluster growing, we need to add new DataNodes to increse the capacity of our cluster. In order to make utilization of every DataNode be in relatively balanced state, usually we use HDFS balancer tool to balance our cluster every time we add new DataNodes.
We have been facing an issue with heterogeneous disk capacity when using HDFS balancer tool. In production cluster, we often have to add new DataNodes with larger disk capacity than previous DNs. Since the original balancer is implemented to balance utilization of every DataNode, the balancer will make every DN's utilization and average utilization of the cluster be within a given threshold.
For example, in a cluster with two DataNodes DN1 and DN2, DN1 has ten disks with 2T capacity, DN2 has ten disks with 10T capacity, the original balancer may make the cluster balanced in the following state:

DataNode	Total Capacity	Used	Remaining	utilization
DN1	20T	18T	2T	90%
DN2	100T	90T	10T	90%

each DN has reached a 90% utilization, in such a case, DN1's capacibility to store new blocks is far less than DN2's. When DN1 is full, all of the new blocks will be written to DN2 and more MR tasks will be scheduled to DN2. As a result, DN2 is overloaded and we can not
make full use of each DN's I/O capacity. In such a case, We wish the balancer could run based on remaining space of every DN. After balancing, every DN's remaining space could be balanced like the following state:

DataNode	Total Capacity	Used	Remaining	utilization
DN1	20T	14T	6T	70%
DN2	100T	94T	6T	94%

In a cluster with balanced remaining space of DN's capacity, every DN will be utilized when writing new blocks to the cluster, on the other hand, every DN's I/O capacity can be utilized when running MR jobs.

Please let me know what you guys think. I will attach a patch if necessary.

Issue Links

• relates to: HDFS-12601 Implement new hdfs balancer's threshold units