I did move out the commit in a separate thread, and using SearchManager refreshing the reader every 2 seconds if required. Did some testing publishing a single document in each test. (there are a combination of addDocument and updateDocument) happening.
Following are the tests we ran (Below are the performance summary for each run). The environment remains the same as said in the earlier thread. 1. 1 threadpool and publishing using 10000 threads. (takes 5ms avg to complete) 2. 5 theadpool and publishing 10000 threads 3. 10 threadpool with 10000 threads publishing 4. 100 threadpool and 100 threads publishing. Summary : ------------- Among all of this, the option 1> seems to perform much better. With 10 ms on an avg. And the bottleneck being addDocument and updateDocument method. This link shows the jprofiler trace of the updateDocument for run 2> 5 threads. https://www.dropbox.com/s/3d41mautf12f373/2014-07-03%2012_15_54-IndexTester%20%5B3%5D%20-%20JProfiler%208.0.7.png Question: --------------- Is there a way to improve the addDocument and updateDocument performance? Analysis Numbers for the test runs --------------------------------------------- 1. 1 threadpool and publishing using 10000 threads. (takes 5ms avg to complete) -------------------------------------------- Min | 3 ms max | 301 ms avg | 5.733 ms -------------------------------------------- >5000 ms | 0 threads >2000 <5000 ms| 0 threads >1000 <2000 ms| 0 threads >500 <1000ms | 0 threads >100 <500ms | 21 threads >20 <50 ms | 61 threads >0 <20 ms | 9888 threads -------------------------------------------- 2. 5 theadpool and publishing 10000 threads -------------------------------------------- Min | 3 ms max | 677 ms avg | 14.8305 ms -------------------------------------------- >5000 ms | 0 threads >2000 <5000 ms| 0 threads >1000 <2000 ms| 0 threads >500 <1000ms | 5 threads >100 <500ms | 105 threads >20 <50 ms | 753 threads >0 <20 ms | 8881 threads -------------------------------------------- 3. 10 threadpool with 10000 threads publishing -------------------------------------------- Min | 3 ms max | 980 ms avg | 31.8305 ms -------------------------------------------- >5000 ms | 0 threads >2000 <5000 ms| 0 threads >1000 <2000 ms| 0 threads >500 <1000ms | 11 threads >100 <500ms | 340 threads >20 <50 ms | 4493 threads >0 <20 ms | 4095 threads -------------------------------------------- 4. 100 threadpool and 100 threads publishing. -------------------------------------------- Min | 109 ms max | 939 ms avg | 651.42 ms -------------------------------------------- >5000 ms | 0 threads >2000 <5000 ms| 0 threads >1000 <2000 ms| 0 threads >500 <1000ms | 74 threads >100 <500ms | 26 threads >20 <50 ms | 0 threads >0 <20 ms | 0 threads -------------------------------------------- -----Original Message----- From: Vitaly Funstein [mailto:vfunst...@gmail.com] Sent: Saturday, June 21, 2014 11:50 AM To: java-user@lucene.apache.org Subject: Re: Concurrent Indexing Hmm, I'm not sure you want to rely on the presence or absence of a particular document in the index to determine the recovery point. It may work for inserts, but not likely for updates or removes. I would look into driving the version numbers from the commiter to the DB, and record them as commit user data for each Lucene index commit. Then on startup or restart of your webapp, you simply grab the version from that data in the reopened Lucene commit point, and query the database for all the records with the version > that number; if you find any, add them to the index, and do one more commit to bring the index up to date. This is probably beyond the scope of your original query, however. On Fri, Jun 20, 2014 at 10:46 PM, Umashanker, Srividhya < srividhya.umashan...@hp.com> wrote: > We do have a way to recover partially with a version number for each > transaction. The same version maintained in lucene as one document. > During startup these numbers define what has to be syncd up. > Unfortunately lucene is used in a webapp, so this happens "only" during a > jetty restart. > > - Vidhya > > > > On 21-Jun-2014, at 11:08 am, "Vitaly Funstein" <vfunst...@gmail.com> > wrote: > > > > This is a better idea than what you had before, but I don't think > > there's any point in doing any commits manually at all unless you > > have a way of detecting and recovering exactly the data that hasn't > > been committed. In other words, what difference does it make whether > > you lost 1 index record or 1M, if you can't determine which records > > were lost and need to reindex everything from the start anyway, to > > ensure consistency between SOR and Lucene? > > > > > > > > > > On Fri, Jun 20, 2014 at 10:20 PM, Umashanker, Srividhya < > > srividhya.umashan...@hp.com> wrote: > > > >> Let me try with the NRT and periodic commit say every 5 mins in a > >> committer thread on need basis. > >> > >> Is there a threshold limit on how long we can go without committing > >> ? I think the buffers get flushed to disk but not to crash proof on disk. > So we > >> should be good on memory. > >> > >> I should also verify if the time taken for commit() is longer when > >> more data piled up to commit. But definitely should be better than > committing > >> for every thread.. > >> > >> Will post back after tests. > >> > >> - Vidhya > >> > >> > >>> On 21-Jun-2014, at 10:28 am, "Vitaly Funstein" > >>> <vfunst...@gmail.com> > >> wrote: > >>> > >>> Hmm, I might have actually given you a slightly incorrect > >>> explanation > wrt > >>> what happens when internal buffers fill up. There will definitely > >>> be a flush of the buffer, and segment files will be written to, > >>> but it's not actually considered a full commit, i.e. an external > >>> reader will not see these changes (yet). The exact details elude > >>> me but there are quite a > few > >>> threads here on what happens during a commit (vs a flush). > >>> However, > when > >>> you call IndexWriter.close() a commit will definitely happen. > >>> > >>> But in any event, if you use an NRT reader to search, then it > >>> shouldn't matter to you when the commit actually takes place. Such > >>> readers also search uncommitted changes as well as those already > >>> on disk. If data durability is not a requirement for you, if i.e. > >>> you can (and probably > >> do) > >>> reindex your data from SOR on startup, then not doing commits > >>> yourself > >> may > >>> be the way to go. Or perhaps you could reduce the amount of data > >>> you > need > >>> to reindex and still call commit() yourself periodically though > >>> not for every write transaction, but maybe introduce some > >>> watermarking logic whereby you detect the highest watermark > >>> committed to Lucene. Then > >> reindex > >>> only the data from the DB from that point onward (meaning only > >> uncommitted > >>> data is lost and needs to be recovered, but you can figure out > >>> exactly where that point is). > >>> > >>> > >>> > >>> On Fri, Jun 20, 2014 at 8:02 PM, Umashanker, Srividhya < > >>> srividhya.umashan...@hp.com> wrote: > >>> > >>>> It is non transactional. We first write the same data to database > >>>> in a transaction and then call writer addDocument. If lucene > >>>> fails we > still > >>>> hold the data to recover. > >>>> > >>>> I can avoid the commit if we use NRT reader. We do need this to > >>>> be searchable immediately. > >>>> > >>>> Another question. I did try removing commit() in each thread and > >>>> wait > >> for > >>>> lucene to auto commit with maxBufferedDocs set to 100 and > >> ramBufferedSize > >>>> set to high values, so docs triggers first. But did not see the > >>>> 1st > 100 > >>>> docs data in lucene even after 500 docs. > >>>> > >>>> Is there a way for me to see when lucene auto commits? > >>>> > >>>> If we tune the auto commit parameters appropriately, do i still > >>>> need > the > >>>> committer thread ? Because it's job is to call commit. Anyway > >>>> add/updateDocument is already done in my writer threads. > >>>> > >>>> Thanks for your time and your suggestions! > >>>> > >>>> - Vidhya > >>>> > >>>> > >>>>> On 21-Jun-2014, at 12:09 am, "Vitaly Funstein" > >>>>> <vfunst...@gmail.com> > >>>> wrote: > >>>>> > >>>>> You could just avoid calling commit() altogether if your > application's > >>>>> semantics allow this (i.e. it's non-transactional in nature). > >>>>> This > way, > >>>>> Lucene will do commits when appropriate, based on the buffering > >> settings > >>>>> you chose. It's generally unnecessary and undesirable to call > >>>>> commit > at > >>>> the > >>>>> end of each write, unless you see to provide strict durability > >> guarantees > >>>>> in your system. > >>>>> > >>>>> If you must acknowledge every write after it's been committed, > >>>>> set > up a > >>>>> single committer thread that does this when there are any work > >>>>> tasks > in > >>>> the > >>>>> queue. Then add to that queue from your writer threads... > >>>>> > >>>>> > >>>>> On Fri, Jun 20, 2014 at 8:47 AM, Umashanker, Srividhya < > >>>>> srividhya.umashan...@hp.com> wrote: > >>>>> > >>>>>> Lucene Experts - > >>>>>> > >>>>>> Recently we upgraded to Lucene 4. We want to make use of > >>>>>> concurrent flushing feature Of Lucene. > >>>>>> > >>>>>> Indexing for us includes certain db operations and writing to > >>>>>> lucene > >>>> ended > >>>>>> by commit. There may be multiple concurrent calls to Indexer > >>>>>> to > >> publish > >>>>>> single/multiple records. > >>>>>> > >>>>>> So far, with older version of lucene, we had our indexing > synchronized > >>>> (1 > >>>>>> thread indexing). > >>>>>> Which means waiting time is more, based on concurrency and > >>>>>> execution > >>>> time. > >>>>>> > >>>>>> We are moving away from the Synchronized indexing. Which is > >>>>>> actually > >> to > >>>>>> cut down the waiting period. Trying to find out if we have to > >>>>>> limit > >> the > >>>>>> number of threads that adds document and commits. > >>>>>> > >>>>>> Below are the tests - to publish just 1000 records with 3 text > fields. > >>>>>> > >>>>>> Java 7 , JVM config : -XX:MaxPermSize=384M > >>>>>> -XX:+HeapDumpOnOutOfMemoryError -Xmx400m -Xms50m > -XX:MaxNewSize=100m > >>>>>> -Xss256k -XX:-UseParallelOldGC -XX:-UseSplitVerifier > >>>>>> -Djsse.enableSNIExtension=false > >>>>>> > >>>>>> IndexConfiguration being default : We also tried with changes > >>>>>> in maxThreadStates,maxBufferedDocs,ramBufferSizeMB - no impact. > >>>>>> > >>>>>> > >>>>>> > >>>>>> Min time in ms > >>>>>> > >>>>>> Max time ms > >>>>>> > >>>>>> Avg time ms > >>>>>> > >>>>>> 1 thread -commit > >>>>>> > >>>>>> 65 > >>>>>> > >>>>>> 267 > >>>>>> > >>>>>> 85 > >>>>>> > >>>>>> 1 thread -updateDocument > >>>>>> > >>>>>> 0 > >>>>>> > >>>>>> 40 > >>>>>> > >>>>>> 1 > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> 6 thread-commit > >>>>>> > >>>>>> 83 > >>>>>> > >>>>>> 1449 > >>>>>> > >>>>>> 552.42 > >>>>>> > >>>>>> 6 thread- updateDocument > >>>>>> > >>>>>> 0 > >>>>>> > >>>>>> 175 > >>>>>> > >>>>>> 1.5 > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> 10 thread -Commit > >>>>>> > >>>>>> 154 > >>>>>> > >>>>>> 2429 > >>>>>> > >>>>>> 874 > >>>>>> > >>>>>> 10 thread- updateDocument > >>>>>> > >>>>>> 0 > >>>>>> > >>>>>> 243 > >>>>>> > >>>>>> 1.9 > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> 20 thread -commit > >>>>>> > >>>>>> 76 > >>>>>> > >>>>>> 4351 > >>>>>> > >>>>>> 1622 > >>>>>> > >>>>>> 20 thread - updateDocument > >>>>>> > >>>>>> 0 > >>>>>> > >>>>>> 326 > >>>>>> > >>>>>> 2.1 > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> > >>>>>> More the threads trying to write to lucene, the updateDocument > >>>>>> and > >>>>>> commit() are becoming bottlenecks. In the above table, 10 and > >>>>>> 20 > >>>> threads > >>>>>> have an average of 1.5 sec for 1000 commits. > >>>>>> > >>>>>> Is there some configuration of suggestions to tune the > >>>>>> performance > of > >>>> the > >>>>>> 2 methods, so that our service performs better, with more > concurrency? > >>>>>> > >>>>>> -vidhya > >>>>>> > >>>>>> > >>>>>> > >>>> > >>>> ----------------------------------------------------------------- > >>>> ---- To unsubscribe, e-mail: > >>>> java-user-unsubscr...@lucene.apache.org > >>>> For additional commands, e-mail: java-user-h...@lucene.apache.org > >>>> > >>>> > >> > >> ------------------------------------------------------------------- > >> -- To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org > >> For additional commands, e-mail: java-user-h...@lucene.apache.org > >> > >> > > --------------------------------------------------------------------- > To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org > For additional commands, e-mail: java-user-h...@lucene.apache.org > >
