Sorry for jumping in late here. What's you compaction queue size over time?
It might be that your IO system just cannot keep up with the load. HBase will buffer data in the memstore, but eventually this data has to make it to disk, then eventually you get a lot of storefiles that need to be compacted. Eventually if the compactions cannot keep up, flushes are blocked, and HBase will block client writes (what else can it do?) The large compaction queue size of a good indicator for that. BTW. HBASE-6974 will add a separate metric for these blocked writes, so that these can be tracker more directly. -- Lars ________________________________ From: Jonathan Bishop <jbishop....@gmail.com> To: user@hbase.apache.org Sent: Sunday, October 14, 2012 8:48 AM Subject: Re: more regionservers does not improve performance Matt, Yes, I did. What I observed is that the map job proceeds about 3-4x faster for a while. But then I observed long pauses partway through the job, and overall run time was only reduced only modestly, way from 50 minutes to 40 minutes. Just to summarize the issue, my mapper jobs seem to scale nicely. This is expected as my dfs block size is small enough to create over 500 tasks, and I have a max of 40 mappers running. But when I include puts to hbase in my job, then I see a 4-6x slowdown which does not respond to an increasing number of regionservers. My current best guess is that there is a network bottleneck in getting the puts produced by the mappers to the appropriate regionservers, as I assume that once the puts are received by the regionservers that they can all operate in parallel without slowing each other down. Again, I am on grid which is used by many others, and the machines in my cluster are not dedicated to my job. I am mainly looking at scalability trends when running with various numbers of regionservers. Jon On Sat, Oct 13, 2012 at 10:37 PM, Matt Corgan <mcor...@hotpads.com> wrote: > Did you try setting put.setWriteToWAL(false) as Bryan suggested? This may > not be what you want in the end, but seeing what happens may help debug. > > Matt > > On Sat, Oct 13, 2012 at 8:58 AM, Jonathan Bishop <jbishop....@gmail.com > >wrote: > > > Suraj, > > > > I bumped my regionservers all the way up to 32g from 8g. They are running > > on 64g and 128g machines on our cluster. Unfortunately, the machines all > > have various states of loading (usually high) from other users. > > > > In ganglia I do not see any swapping, but that has been known to happen > > from time to time. > > > > Thanks for your help - I'll take a look at your links. > > > > Jon > > > > On Fri, Oct 12, 2012 at 7:30 PM, Suraj Varma <svarma...@gmail.com> > wrote: > > > > > Hi Jonathan: > > > What specific metric on ganglia did you notice for "IO is spiking"? Is > > > it your disk IO? Is your disk swapping? Do you see cpu iowait spikes? > > > > > > I see you have given 8g to the RegionServer ... how much RAM is > > > available total on that node? What heap are the individual mappers & > > > DN set to run on (i.e. check whether you are overallocated on heap > > > when the _mappers_ run ... causing disk swapping ... leading to IO?). > > > > > > There can be multiple causes ... so, you may need to look at ganglia > > > stats and narrow the bottleneck down as described in > > > http://hbase.apache.org/book/casestudies.perftroub.html > > > > > > Here's a good reference for all the memstore related tweaks you can > > > try (and also to understand what each configuration means): > > > > http://blog.sematext.com/2012/07/16/hbase-memstore-what-you-should-know/ > > > > > > Also, provide more details on your schema (CFs, row size), Put sizes, > > > etc as well to see if that triggers an idea from the list. > > > --S > > > > > > > > > On Fri, Oct 12, 2012 at 12:46 PM, Bryan Beaudreault > > > <bbeaudrea...@hubspot.com> wrote: > > > > I recommend turning on debug logging on your region servers. You may > > > need > > > > to tune down certain packages back to info, because there are a few > > > spammy > > > > ones, but overall it helps. > > > > > > > > You should see messages such as "12/10/09 14:22:57 INFO > > > > regionserver.HRegion: Blocking updates for 'IPC Server handler 41 on > > > 60020' > > > > on region XXX: memstore size 256.0m is >= than blocking 256.0m size". > > As > > > > you can see, this is an INFO anyway so you should be able to see it > now > > > if > > > > it is happening. > > > > > > > > You can try upping the number of IPC handlers and the memstore flush > > > > threshold. Also, maybe you are bottlenecked by the WAL. Try doing > > > > put.setWriteToWAL(false), just to see if it increases performance. > If > > so > > > > and you want to be a bit more safe with regard to the wal, you can > try > > > > turning on deferred flush on your table. I don't really know how to > > > > increase performance of the wal aside from that, if this does seem to > > > have > > > > an affect. > > > > > > > > > > > > > > > > On Fri, Oct 12, 2012 at 3:15 PM, Jonathan Bishop < > > jbishop....@gmail.com > > > >wrote: > > > > > > > >> Kevin, > > > >> > > > >> Sorry, I am fairly new to HBase. Can you be specific about what > > > settings I > > > >> can change, and also where they are specified? > > > >> > > > >> Pretty sure I am not hotspotting, and increasing memstore does not > > seem > > > to > > > >> have any effect. > > > >> > > > >> I do not seen any messages in my regionserver logs concerning > > blocking. > > > >> > > > >> I am suspecting that I am hitting some limit in our grid, but would > > > like to > > > >> know where that limit is being imposed. > > > >> > > > >> Jon > > > >> > > > >> On Fri, Oct 12, 2012 at 6:44 AM, Kevin O'dell < > > kevin.od...@cloudera.com > > > >> >wrote: > > > >> > > > >> > Jonathan, > > > >> > > > > >> > Lets take a deeper look here. > > > >> > > > > >> > What is your memstore set at for the table/CF in question? Lets > > > compare > > > >> > that value with the flush size you are seeing for your regions. > If > > > they > > > >> > are really small flushes is it all to the same region? If so that > > is > > > >> going > > > >> > to be schema issues. If they are full flushes you can up your > > > memstore > > > >> > assuming you have the heap to cover it. If they are smaller > flushes > > > but > > > >> to > > > >> > different regions you most likely are suffering from global limit > > > >> pressure > > > >> > and flushing too soon. > > > >> > > > > >> > Are you flushing prematurely due to HLogs rolling? Take a look > for > > > too > > > >> > many hlogs and look at the flushes. It may benefit you to raise > > that > > > >> > value. > > > >> > > > > >> > Are you blocking? As Suraj was saying you may be blocking in > > 90second > > > >> > blocks. Check the RS logs for those messages as well and then > > Suraj's > > > >> > advice. > > > >> > > > > >> > This is where I would start to optimize your write path. I hope > the > > > >> above > > > >> > helps. > > > >> > > > > >> > On Fri, Oct 12, 2012 at 3:34 AM, Suraj Varma <svarma...@gmail.com > > > > > >> wrote: > > > >> > > > > >> > > What have you configured your hbase.hstore.blockingStoreFiles > and > > > >> > > hbase.hregion.memstore.block.multiplier? Both of these block > > updates > > > >> > > when the limit is hit. Try increasing these to say 20 and 4 from > > the > > > >> > > default 7 and 2 and see if it helps. > > > >> > > > > > >> > > If this still doesn't help, see if you can set up ganglia to > get a > > > >> > > better insight into what is bottlenecking. > > > >> > > --Suraj > > > >> > > > > > >> > > > > > >> > > > > > >> > > On Thu, Oct 11, 2012 at 11:47 PM, Pankaj Misra > > > >> > > <pankaj.mi...@impetus.co.in> wrote: > > > >> > > > OK, Looks like I missed out reading that part in your original > > > mail. > > > >> > Did > > > >> > > you try some of the compaction tweaks and configurations as > > > explained > > > >> in > > > >> > > the following link for your data? > > > >> > > > http://hbase.apache.org/book/regions.arch.html#compaction > > > >> > > > > > > >> > > > > > > >> > > > Also, how much data are your putting into the regions, and how > > > big is > > > >> > > one region at the end of data ingestion? > > > >> > > > > > > >> > > > Thanks and Regards > > > >> > > > Pankaj Misra > > > >> > > > > > > >> > > > -----Original Message----- > > > >> > > > From: Jonathan Bishop [mailto:jbishop....@gmail.com] > > > >> > > > Sent: Friday, October 12, 2012 12:04 PM > > > >> > > > To: user@hbase.apache.org > > > >> > > > Subject: RE: more regionservers does not improve performance > > > >> > > > > > > >> > > > Pankaj, > > > >> > > > > > > >> > > > Thanks for the reply. > > > >> > > > > > > >> > > > Actually, I am using MD5 hashing to evenly spread the keys > among > > > the > > > >> > > splits, so I don’t believe there is any hotspot. In fact, when I > > > >> monitory > > > >> > > the web UI for HBase I see a very even load on all the > > > regionservers. > > > >> > > > > > > >> > > > Jon > > > >> > > > > > > >> > > > Sent from my Windows 8 PC < > > > >> > http://windows.microsoft.com/consumer-preview > > > >> > > > > > > >> > > > > > > >> > > > *From:* Pankaj Misra <pankaj.mi...@impetus.co.in> > > > >> > > > *Sent:* Thursday, October 11, 2012 8:24:32 PM > > > >> > > > *To:* user@hbase.apache.org > > > >> > > > *Subject:* RE: more regionservers does not improve performance > > > >> > > > > > > >> > > > Hi Jonathan, > > > >> > > > > > > >> > > > What seems to me is that, while doing the split across all 40 > > > >> mappers, > > > >> > > the keys are not randomized enough to leverage multiple regions > > and > > > the > > > >> > > pre-split strategy. This may be happening because all the 40 > > mappers > > > >> may > > > >> > be > > > >> > > trying to write onto a single region for sometime, making it a > HOT > > > >> > region, > > > >> > > till the key falls into another region, and then the other > region > > > >> > becomes > > > >> > > a HOT region hence you may seeing a high impact of compaction > > cycles > > > >> > > reducing your throughput. > > > >> > > > > > > >> > > > Are the keys incremental? Are the keys randomized enough > across > > > the > > > >> > > splits? > > > >> > > > > > > >> > > > Ideally when all 40 mappers are running you should see all the > > > >> regions > > > >> > > being filled up in parallel for maximum throughput. Hope it > helps. > > > >> > > > > > > >> > > > Thanks and Regards > > > >> > > > Pankaj Misra > > > >> > > > > > > >> > > > > > > >> > > > ________________________________________ > > > >> > > > From: Jonathan Bishop [jbishop....@gmail.com] > > > >> > > > Sent: Friday, October 12, 2012 5:38 AM > > > >> > > > To: user@hbase.apache.org > > > >> > > > Subject: more regionservers does not improve performance > > > >> > > > > > > >> > > > Hi, > > > >> > > > > > > >> > > > I am running a MR job with 40 simultaneous mappers, each of > > which > > > >> does > > > >> > > puts to HBase. I have ganged up the puts into groups of 1000 > (this > > > >> seems > > > >> > to > > > >> > > help quite a bit) and also made sure that the table is pre-split > > > into > > > >> 100 > > > >> > > regions, and that the row keys are randomized using MD5 hashing. > > > >> > > > > > > >> > > > My cluster size is 10, and I am allowing 4 mappers per > > > tasktracker. > > > >> > > > > > > >> > > > In my MR job I know that the mappers are able to generate puts > > > much > > > >> > > faster than the puts can be handled in hbase. In other words if > I > > > let > > > >> the > > > >> > > mappers run without doing hbase puts then everything scales as > you > > > >> would > > > >> > > expect with the number of mappers created. It is the hbase puts > > > which > > > >> > seem > > > >> > > to be the bottleneck. > > > >> > > > > > > >> > > > What is strange is that I do not get much run time improvement > > by > > > >> > > increasing the number regionservers beyond about 4. Indeed, it > > seems > > > >> that > > > >> > > the system runs slower with 8 regionservers than with 4. > > > >> > > > > > > >> > > > I have added the following in hbase-env.sh hoping this would > > > help... > > > >> > > (from the book HBase in Action) > > > >> > > > > > > >> > > > export HBASE_OPTS="-Xmx8g" > > > >> > > > export HBASE_REGIONSERVER_OPTS="-Xmx8g -Xms8g -Xmn128m > > > >> -XX:+UseParNewGC > > > >> > > -XX:+UseConcMarkSweepGC -XX:CMSInitiatingOccupancyFraction=70" > > > >> > > > > > > >> > > > # Uncomment below to enable java garbage collection logging in > > the > > > >> .out > > > >> > > file. > > > >> > > > export HBASE_OPTS="${HBASE_OPTS} -verbose:gc > -XX:+PrintGCDetails > > > >> > > -XX:+PrintGCDateStamps -Xloggc:${HBASE_HOME}/logs/gc-hbase.log" > > > >> > > > > > > >> > > > Monitoring hbase through the web ui I see that there are > pauses > > > for > > > >> > > flushing, which seems to run pretty quickly, and for compacting, > > > which > > > >> > > seems to take somewhat longer. > > > >> > > > > > > >> > > > Any advice for making this run faster would be greatly > > > appreciated. > > > >> > > > Currently I am looking into installing Ganglia to better > > monitory > > > my > > > >> > > cluster, but yet to have that running. > > > >> > > > > > > >> > > > I suspect an I/O issue as the regionservers do not seem > terribly > > > >> > loaded. > > > >> > > > > > > >> > > > Thanks, > > > >> > > > > > > >> > > > Jon > > > >> > > > > > > >> > > > ________________________________ > > > >> > > > > > > >> > > > Impetus Ranked in the Top 50 India’s Best Companies to Work > For > > > 2012. > > > >> > > > > > > >> > > > Impetus webcast ‘Designing a Test Automation Framework for > > > >> Multi-vendor > > > >> > > Interoperable Systems’ available at http://lf1.me/0E/. > > > >> > > > > > > >> > > > > > > >> > > > NOTE: This message may contain information that is > confidential, > > > >> > > proprietary, privileged or otherwise protected by law. The > message > > > is > > > >> > > intended solely for the named addressee. If received in error, > > > please > > > >> > > destroy and notify the sender. Any use of this email is > prohibited > > > when > > > >> > > received in error. Impetus does not represent, warrant and/or > > > >> guarantee, > > > >> > > that the integrity of this communication has been maintained nor > > > that > > > >> the > > > >> > > communication is free of errors, virus, interception or > > > interference. > > > >> > > > > > > >> > > > ________________________________ > > > >> > > > > > > >> > > > Impetus Ranked in the Top 50 India’s Best Companies to Work > For > > > 2012. > > > >> > > > > > > >> > > > Impetus webcast ‘Designing a Test Automation Framework for > > > >> Multi-vendor > > > >> > > Interoperable Systems’ available at http://lf1.me/0E/. > > > >> > > > > > > >> > > > > > > >> > > > NOTE: This message may contain information that is > confidential, > > > >> > > proprietary, privileged or otherwise protected by law. The > message > > > is > > > >> > > intended solely for the named addressee. If received in error, > > > please > > > >> > > destroy and notify the sender. Any use of this email is > prohibited > > > when > > > >> > > received in error. Impetus does not represent, warrant and/or > > > >> guarantee, > > > >> > > that the integrity of this communication has been maintained nor > > > that > > > >> the > > > >> > > communication is free of errors, virus, interception or > > > interference. > > > >> > > > > > >> > > > > >> > > > > >> > > > > >> > -- > > > >> > Kevin O'Dell > > > >> > Customer Operations Engineer, Cloudera > > > >> > > > > >> > > > > > >