hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Tim Robertson <timrobertson...@gmail.com>
Subject Re: PerformanceEvaluation scan - how to read the results?
Date Thu, 26 Jan 2012 09:36:42 GMT
Hey stack

>> This gave me 32 regions across 2 of our 3 region servers (we have HDFS
>> across 17 nodes but only machines running 3 RS).
> The balancer ran?  I'd think it'd balance the regions across the three
> servers.  Something stuck in transition stopping the balancer running
> (See master log).

The cluster is balanced on the whole, but was only using 2 regions for
the TestTable:
12/01/26 09:43:41 INFO master.LoadBalancer: Skipping load balancing.
servers=3 regions=1154 average=384.66666 mostloaded=385

>> And then the following to scan:
>> $HADOOP_HOME/bin/hadoop org.apache.hadoop.hbase.PerformanceEvaluation scan 5
> So, sounds like we're going against two of the three servers only.
>> The output of the scan is:
>> 12/01/25 15:11:02 INFO mapred.JobClient:     ROWS=5242850
>> 12/01/25 15:11:02 INFO mapred.JobClient:     ELAPSED_TIME=1624832
>> (job took 52 secs in reality)
>> Can anyone elaborate on how I am meant to interpret these numbers
>> please?  Looks like 3.2 rows per <timeunit>
> Your MR job scanned 5M rows.  It looks like you had 5 clients so you
> should have had 5 mappers running.  The ELAPSED_TIME is supposed to be
> the sum of the elapsed time of all mappers.  The above looks way wrong
> to me.

No, we got 50 Mappers...
  12/01/26 10:32:10 INFO hbase.PerformanceEvaluation: Total # of splits: 50

Does this indicate something suspicious perhaps?

>> [I am trying to benchmark because our real data of 340M rows (215G on
>> HDFS) takes 60 mins to scan which seems a lot]
> Three servers?  Scanning in sequence?  What rate you seeing per server
> Tim?  What kind of servers (I think you've posted your profile the
> list before but ... it was a while back (smile)).  What size the rows
> being returned?

So we have a really unbalanced cluster with 3 classes of machine.  17
nodes are running HDFS and MR, and 3 of them running HBase.
The RS are running on:
- 2x Intel(R) Xeon(R) CPU E5630 @ 2.53GHz (quad)
- 6x250G SATA 5.4K
- 24GB

A full scan of the big tables (running through the API, or using Hive)
runs in about 1hr, which basic maths suggests to me about 20k rows /
sec / region (approx. 4kb per row)
Because for these tables, we have the probability of non local data
since the HDFS balancer runs, I want to sanitize things using the PE,
before we go and buy new machines.

For the PE test I've checked the TestTable is local to the RS.  I
think it is best to try and confirm the PE is performing as expected
before looking at our data.


View raw message