How the MapR balance data within the whole cluster?

Question

I setup a 6-nodes cluster for 1 week, and I find the data seems not well balanced among all disks in the cluster.

Take today's data for instance:

master: 18% of 652.0GB in use for 1 MapR-FS disk(s) (used: 120.0 GB)

slave1: 9% of 942.0GB in use for 1 MapR-FS disk(s) (used: 88.0 GB)

slave2: 8% of 946.0GB in use for 1 MapR-FS disk(s) (used: 71.0 GB)

slave3: 8% of 946.0GB in use for 1 MapR-FS disk(s) (used: 75.0 GB)

slave4: 7% of 952.0GB in use for 1 MapR-FS disk(s) (used: 71.0 GB)

slave5: 8% of 942.0GB in use for 1 MapR-FS disk(s) (used: 73.0 GB)

Data seems prefer to the master-node, but why?

And what is the strategy for the data-balancing?

How can I balance data in my cluster?

Accepted Answer

The MapR system balances automatically by moving data from nodes that are more full than the cluster average. A node that is within +/-10% of the cluster-average is considered to be within average fullness.

Additionally, by default the balancer doesn't move data from a node unless it is atleast 70% full. This behavior can be changed by modifying the config variable "cldb.balancer.disk.threshold.percentage".

In the instance that you pasted, all of the nodes have disk-utilization within 10% of the average. And, all of them are < 70%. So, the balancer would take no action.

By master, I'm assuming you mean the CLDB node. MapR tries hard to keep the first copy of every write local - if this causes an anamoly that causes some nodes to have excessive space utilization compared to the rest of the cluster (+/- 10% of cluster-avg), the balancer will fix it.

Please look at http://www.mapr.com/doc/display/MapR/Balancers for additional details regarding the balancer.

Answer 2

0	Is the disk balancer running on your cluster? Cut and paste the output of: /opt/mapr/bin/maprcli config load -json \| grep "cldb.balancer" Also, how many volumes do you have? link answered 21 Sep '11, 20:43 Vivek 41 accept rate: 33%

Answer 3

hadoop@lord:~$ /opt/mapr/bin/maprcli config load -json | grep "cldb.balancer"

                    "cldb.balancer.disk.max.switches.in.nodes.percentage":"10",

                    "cldb.balancer.disk.paused":"0",

                    "cldb.balancer.disk.sleep.interval.sec":"120",

                    "cldb.balancer.disk.threshold.percentage":"70",

                    "cldb.balancer.logging":"0",

                    "cldb.balancer.role.max.switches.in.nodes.percentage":"10",

                    "cldb.balancer.role.paused":"0",

                    "cldb.balancer.role.sleep.interval.sec":"900",

                    "cldb.balancer.startup.interval.sec":"1800",

I have 18 system-volumes and 11 user-data-volumes in my cluster.

The data in system-volumes is a little, most data is in the user-data-volume:

vol1: 0

vol2: 421.0 MB

vol3: 450.0 MB

vol4: 13.1 GB

vol5: 20.9 GB

vol6: 23.7 GB

vol7: 30.5 GB

vol8: 34.5 GB

vol9: 36.3 GB

vol10 43.6 GB

vol11: 61.3 GB

How the MapR balance data within the whole cluster?

Follow this question

Related questions