[
https://issues.apache.org/jira/browse/HBASE-13965?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14617464#comment-14617464
]
Ted Yu commented on HBASE-13965:
--------------------------------
Loaded patch v5 on a small cluster and obtained the following:
{code}
}, {
"name" : "Hadoop:service=HBase,name=Master,sub=Balancer",
"modelerType" : "Master,sub=Balancer",
"tag.Context" : "master",
"tag.Hostname" : "cn013.l42scl.hortonworks.com",
"IntegrationTestBigLinkedList_StoreFileCostFunction" : 3.262317387568032,
"IntegrationTestBigLinkedList_LocalityCostFunction" : 2.473958333333334,
"IntegrationTestBigLinkedList_TableSkewCostFunction" : 5.605468749999999,
"IntegrationTestBigLinkedList_Overall" : 35.09174447090137,
"IntegrationTestBigLinkedList_WriteRequestCostFunction" : 0.0,
"IntegrationTestBigLinkedList_RegionCountSkewCostFunction" : 0.0,
"IntegrationTestBigLinkedList_ReadRequestCostFunction" : 5.0,
"IntegrationTestBigLinkedList_MemstoreSizeCostFunction" : 0.0,
"IntegrationTestBigLinkedList_RegionReplicaHostCostFunction" : 0.0,
"IntegrationTestBigLinkedList_RegionReplicaRackCostFunction" : 0.0,
"IntegrationTestBigLinkedList_MoveCostFunction" : 18.75,
{code}
Do you think it makes sense to expose each cost (other than Overall) as
percentage ?
This way, it is easier for user to figure out which cost is the dominant factor.
> Stochastic Load Balancer JMX Metrics
> ------------------------------------
>
> Key: HBASE-13965
> URL: https://issues.apache.org/jira/browse/HBASE-13965
> Project: HBase
> Issue Type: Improvement
> Components: Balancer, metrics
> Reporter: Lei Chen
> Assignee: Lei Chen
> Attachments: HBASE-13965-v3.patch, HBASE-13965-v4.patch,
> HBASE-13965-v5.patch, HBASE-13965_v2.patch, HBase-13965-v1.patch,
> stochasticloadbalancerclasses_v2.png
>
>
> Today’s default HBase load balancer (the Stochastic load balancer) is cost
> function based. The cost function weights are tunable but no visibility into
> those cost function results is directly provided.
> A driving example is a cluster we have been tuning which has skewed rack size
> (one rack has half the nodes of the other few racks). We are tuning the
> cluster for uniform response time from all region servers with the ability to
> tolerate a rack failure. Balancing LocalityCost, RegionReplicaRack Cost and
> RegionCountSkew Cost is difficult without a way to attribute each cost
> function’s contribution to overall cost.
> What this jira proposes is to provide visibility via JMX into each cost
> function of the stochastic load balancer, as well as the overall cost of the
> balancing plan.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
