Hi Pere,

  I have a few thoughts here:

1.) I would make sure you have these settings - -XX:+UseParNewGC
-XX:+UseConcMarkSweepGC -XX:-CMSConcurrentMTEnabled
-XX:CMSInitiatingOccupancyFraction=70 -XX:+CMSParallelRemarkEnabled

2.) Does this happen with a lower scanner.caching?

3.) Did it go away when you turned the block cache off?

4.) Do you have MSLab turned on?

On Thu, Nov 13, 2014 at 10:12 PM, Dhaval Shah <prince_mithi...@yahoo.co.in>
wrote:

> You can do scan.setCacheBlocks (false) To disable scanner caching on map
> reduce scans. Also use parallel GC for new generation.  That will help
> reduce stop the world pauses with cms
>
>
> Sent from my T-Mobile 4G LTE Device
>
>
> -------- Original message --------
> From: Pere Kyle <p...@whisper.sh>
> Date:11/13/2014  8:54 PM  (GMT-05:00)
> To: user@hbase.apache.org
> Cc:
> Subject: Avoid GC Pauses on Scan MapReduces
>
> Hi there,
>
> Recently I have been experiencing instability when scanning our HBASE
> cluster. The table we are trying to scan is 1.5B records 1TB, we have 12GB
> heap and 17 servers. Our GC options are as so:
> -XX:OnOutOfMemoryError=kill -9 %p -Xmx12000m -XX:+UseConcMarkSweepGC
> -Xmx12g -Xmx12g
>
> The errors I am seeing in the logs:
> 2014-11-14 00:00:32,656 WARN org.apache.hadoop.hbase.util.Sleeper
> (regionserver60020): We slept 17355ms instead of 3000ms, this is likely due
> to a long garbage collecting pause and it's usually bad, see
> http://hbase.apache.org/book.html#trouble.rs.runtime.zkexpired
> 2014-11-14 00:00:32,663 WARN org.apache.hadoop.hbase.util.Sleeper
> (regionserver60020.periodicFlusher): We slept 24734ms instead of 10000ms,
> this is likely due to a long garbage collecting pause and it's usually bad,
> see http://hbase.apache.org/book.html#trouble.rs.runtime.zkexpired
>
> We are using hbase.client.scanner.caching=1000. I suspect this may be a
> block cache issue. My question is if/how to disable the block cache for the
> scan queries? This is taking out writes and causing instability on the
> cluster.
>
> Thanks,
> Pere
>



-- 
Kevin O'Dell
Systems Engineer, Cloudera

Reply via email to