Duo Zhang created HBASE-25832:
---------------------------------

             Summary: Performance improvements for StochasticLoadBalancer
                 Key: HBASE-25832
                 URL: https://issues.apache.org/jira/browse/HBASE-25832
             Project: HBase
          Issue Type: Improvement
          Components: Balancer, Performance
            Reporter: Duo Zhang


This is what I posted in the slack channel.

{quote}
When refactoring balancer related code for moving them out to hbase-balancer 
module,  I found that we will create a OffPeakHours instance in 
MoveCostFunction.cost every time. So I wrote a simple JMH test to test its 
performance.

https://github.com/Apache9/hbase-jmh-test/blob/master/src/main/java/org/apache/hadoop/hbase/master/balancer/MoveCostFunctionBenchmark.java

For the original version,  the result is
Benchmark                                            Mode  Samples    Score  
Score error  Units
o.a.h.h.m.b.MoveCostFunctionBenchmark.computeCost    avgt       10  242.312     
   3.853  ns/op
12:06
If I initialized OffPeakHours in constructor, then the result is
Benchmark                                            Mode  Samples   Score  
Score error  Units
o.a.h.h.m.b.MoveCostFunctionBenchmark.computeCost    avgt       10  81.186      
  0.720  ns/op
12:06
If I also eliminate the conf.getFloat in the cost method(by storing them to a 
field member in constructor, the result is
Benchmark                                            Mode  Samples  Score  
Score error  Units
o.a.h.h.m.b.MoveCostFunctionBenchmark.computeCost    avgt       10  4.419       
 0.035  ns/op
12:08
I think on a large cluster, the performance for balancer is a problem, so maybe 
we should try to avoid creating too many objects or calling conf.getXXX in the 
critical path? As we will call StochasticLoadBalancer.computeCost in every step 
when balancing cluster...
{quote}

For large clusters, the performance of StochasticLoadBalancer will be very 
important, so I think we need to treat this problem seriously and try to 
improve.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

Reply via email to