[jira] [Commented] (LUCENE-3425) NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control
[ https://issues.apache.org/jira/browse/LUCENE-3425?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13767770#comment-13767770 ] Michael McCandless commented on LUCENE-3425: OK, thanks for the explanation; now I understand AverageMergePolicy's purpose, and it makes sense. It's ironic that a fully optimized index is the worst thing you could do when searching segments concurrently ... But, I still don't understand why AverageMergePolicy is not merging the little segments from NRTCachingDir. Do you tell it to target a maximum number of segments in the index? If so, once the index is large enough, it seems like that'd force the small segments to be merged. Maybe, you could also tell it a minimum size of the segments, so that it would merge away any segments still held in NRTCachingDir? NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control Key: LUCENE-3425 URL: https://issues.apache.org/jira/browse/LUCENE-3425 Project: Lucene - Core Issue Type: Improvement Components: core/index Affects Versions: 3.4, 4.0-ALPHA Reporter: Shay Banon Fix For: 5.0, 4.5 A discussion on IRC raised several improvements that can be made to NRT caching dir. Some of the problems it currently has are: 1. Not explicitly controlling the memory usage, which can result in overusing memory (for example, large new segments being committed because refreshing is too far behind). 2. Heap fragmentation because of constant allocation of (probably promoted to old gen) byte buffers. 3. Not being able to control the memory usage across indices for multi index usage within a single JVM. A suggested solution (which still needs to be ironed out) is to have a BufferAllocator that controls allocation of byte[], and allow to return unused byte[] to it. It will have a cap on the size of memory it allows to be allocated. The NRT caching dir will use the allocator, which can either be provided (for usage across several indices) or created internally. The caching dir will also create a wrapped IndexOutput, that will flush to the main dir if the allocator can no longer provide byte[] (exhausted). When a file is flushed from the cache to the main directory, it will return all the currently allocated byte[] to the BufferAllocator to be reused by other files. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-3425) NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control
[ https://issues.apache.org/jira/browse/LUCENE-3425?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13766169#comment-13766169 ] caviler commented on LUCENE-3425: - Yes, i changed the default MergePolicy to AverageMergePolicy, i closed old reader after reopenIfChanged, before upgrade to Lucene 4.4.0, we use this AverageMergePolicy is work well, recently we upgrade to Lucene 4.4.0 and modified this AverageMergePolicy to adapt version 4.4.0, after this, in testing(in very short time we add a lot of documents), we noticed the small segments is so many(more than 40,000), and those small segments seems not have chance to be merged? What difference between version 3.6.2 and 4.4.0 behavior in merge process??? our index files about 72G, split to 148 segment by AverageMergePolicy, every segment about 500M size. {code} public class AverageMergePolicy extends MergePolicy { /** * Default noCFSRatio. If a merge's size is = 10% of the index, then we * disable compound file for it. * * @see #setNoCFSRatio */ public static final double DEFAULT_NO_CFS_RATIO = 0.1; private long maxSegmentSizeMB = 100L;// protected double noCFSRatio = DEFAULT_NO_CFS_RATIO; private booleanpartialExpunge = false; // protected boolean useCompoundFile = true; public AverageMergePolicy() { } @Override public void close() { } @Override public MergeSpecification findForcedDeletesMerges(final SegmentInfos infos) throws CorruptIndexException, IOException { try { // SegmentInfoPerCommit best = null; final int numSegs = infos.size(); for (int i = 0; i numSegs; i++) { final SegmentInfoPerCommit info = infos.info(i); if (info.hasDeletions()) { if (null == best || info.getDelCount() best.getDelCount()) { best = info; } } } final CollectionSegmentInfoPerCommit mergingSegments = writer.get().getMergingSegments(); if (mergingSegments.contains(best)) { return null; // skip merging segment } final MergeSpecification spec = new MergeSpecification(); if (null != best) { spec.add(new OneMerge(Collections.singletonList(best))); } return spec; } catch (final Throwable e) { e.printStackTrace(); return null; } } @Override public MergeSpecification findForcedMerges(final SegmentInfos infos, final int maxNumSegments, final MapSegmentInfoPerCommit, Boolean segmentsToMerge) throws IOException { return findMerges(MergeTrigger.EXPLICIT, infos); } @Override public MergeSpecification findMerges(final MergeTrigger mergeTrigger, final SegmentInfos infos) throws IOException { // partialExpunge = false; // // partialExpunge = true; final long maxSegSize = maxSegmentSizeMB * 1024L * 1024L; // long bestSegSize = maxSegSize; // try { final int numSegs = infos.size(); int numBestSegs = numSegs; { // SegmentInfoPerCommit info; long totalSegSize = 0; // compute the total size of segments for (int i = 0; i numSegs; i++) { info = infos.info(i); final long size = size(info); totalSegSize += size; } numBestSegs = (int) ((totalSegSize + bestSegSize - 1) / bestSegSize); // bestSegSize = (numBestSegs == 0) ? totalSegSize : (totalSegSize + maxSegSize - 1) / numBestSegs; // if (bestSegSize maxSegSize) { bestSegSize = maxSegSize; // numBestSegs = (int) ((totalSegSize + maxSegSize - 1) / bestSegSize); } } MergeSpecification spec = findOneMerge(infos, bestSegSize); //int branch = 0; if (null == spec partialExpunge) { //branch = 1; // final OneMerge expunge = findOneSegmentToExpunge(infos, 0); if (expunge != null) { spec = new MergeSpecification(); spec.add(expunge); } } // MergeLogger.collect(branch, spec);
[jira] [Commented] (LUCENE-3425) NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control
[ https://issues.apache.org/jira/browse/LUCENE-3425?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13766430#comment-13766430 ] caviler commented on LUCENE-3425: - Because inside IndexSearcher.search method, it use thread pool to execute Segment.search in every segments, so, if some segment is too big, it will causes this big segment's searcher too slow, eventually entire search method will too slow. example, we have 2G index files. = use AverageMergePolicy, IndexSearcher.search spent time = 1s segment sizesegment.search spent time 1 500M 1s 2 500M 1s 3 500M 1s 4 500M 1s = use other MergePolicy, IndexSearcher.search spent time = 5s segment size segment.search spent time 1 2000M5s = Why not use LogByteSizeMergePolicy but AverageMergePolicy? Because: 1. I want every semgent as small as possible! 2. I want semgent as more as possible! we don't known how big of one segment size when entire index is growing, so can't use LogByteSizeMergePolicy. if use LogByteSizeMergePolicy and setMaxMergeMB(200M): index = 200M, LogByteSizeMergePolicy = 1 segment(per 200M), AverageMergePolicy = 4 segments(per 50M) index = 2000M, LogByteSizeMergePolicy = 10 segment(per 200M), AverageMergePolicy = 4 segments(per 500M) NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control Key: LUCENE-3425 URL: https://issues.apache.org/jira/browse/LUCENE-3425 Project: Lucene - Core Issue Type: Improvement Components: core/index Affects Versions: 3.4, 4.0-ALPHA Reporter: Shay Banon Fix For: 5.0, 4.5 A discussion on IRC raised several improvements that can be made to NRT caching dir. Some of the problems it currently has are: 1. Not explicitly controlling the memory usage, which can result in overusing memory (for example, large new segments being committed because refreshing is too far behind). 2. Heap fragmentation because of constant allocation of (probably promoted to old gen) byte buffers. 3. Not being able to control the memory usage across indices for multi index usage within a single JVM. A suggested solution (which still needs to be ironed out) is to have a BufferAllocator that controls allocation of byte[], and allow to return unused byte[] to it. It will have a cap on the size of memory it allows to be allocated. The NRT caching dir will use the allocator, which can either be provided (for usage across several indices) or created internally. The caching dir will also create a wrapped IndexOutput, that will flush to the main dir if the allocator can no longer provide byte[] (exhausted). When a file is flushed from the cache to the main directory, it will return all the currently allocated byte[] to the BufferAllocator to be reused by other files. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-3425) NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control
[ https://issues.apache.org/jira/browse/LUCENE-3425?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13766406#comment-13766406 ] Michael McCandless commented on LUCENE-3425: Hmm, I don't understand what AverageMergePolicy is doing? Can you describe its purpose at a high level? Somehow it's failing to merge those 40,000 small segments? And offhand I don't know what changed between 4.3 and 4.4 that would cause AverageMergePolicy to stop merging small segments. Maybe turn on IndexWriter's infoStream and watch which merges are being selected. NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control Key: LUCENE-3425 URL: https://issues.apache.org/jira/browse/LUCENE-3425 Project: Lucene - Core Issue Type: Improvement Components: core/index Affects Versions: 3.4, 4.0-ALPHA Reporter: Shay Banon Fix For: 5.0, 4.5 A discussion on IRC raised several improvements that can be made to NRT caching dir. Some of the problems it currently has are: 1. Not explicitly controlling the memory usage, which can result in overusing memory (for example, large new segments being committed because refreshing is too far behind). 2. Heap fragmentation because of constant allocation of (probably promoted to old gen) byte buffers. 3. Not being able to control the memory usage across indices for multi index usage within a single JVM. A suggested solution (which still needs to be ironed out) is to have a BufferAllocator that controls allocation of byte[], and allow to return unused byte[] to it. It will have a cap on the size of memory it allows to be allocated. The NRT caching dir will use the allocator, which can either be provided (for usage across several indices) or created internally. The caching dir will also create a wrapped IndexOutput, that will flush to the main dir if the allocator can no longer provide byte[] (exhausted). When a file is flushed from the cache to the main directory, it will return all the currently allocated byte[] to the BufferAllocator to be reused by other files. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-3425) NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control
[ https://issues.apache.org/jira/browse/LUCENE-3425?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13765217#comment-13765217 ] caviler commented on LUCENE-3425: - Actually, in my case, using NRTCachingDir always causes OutOfMemory error! after it happened, we analysed a gc.hprof file by Eclipse MAT, we believe that: core segment reader will hold a big size of RAMFile instance that created by NRTCachingDir.createOutput, if core segment reader not merged or closed, those RAMFile instances will use more and more buffer, but never have chance to uncache from NRTCachingDir instance, so, when all segment reader allocated buffer size is bigger than we have, it will causes OutOfMemory error! NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control Key: LUCENE-3425 URL: https://issues.apache.org/jira/browse/LUCENE-3425 Project: Lucene - Core Issue Type: Improvement Components: core/index Affects Versions: 3.4, 4.0-ALPHA Reporter: Shay Banon Fix For: 5.0, 4.5 A discussion on IRC raised several improvements that can be made to NRT caching dir. Some of the problems it currently has are: 1. Not explicitly controlling the memory usage, which can result in overusing memory (for example, large new segments being committed because refreshing is too far behind). 2. Heap fragmentation because of constant allocation of (probably promoted to old gen) byte buffers. 3. Not being able to control the memory usage across indices for multi index usage within a single JVM. A suggested solution (which still needs to be ironed out) is to have a BufferAllocator that controls allocation of byte[], and allow to return unused byte[] to it. It will have a cap on the size of memory it allows to be allocated. The NRT caching dir will use the allocator, which can either be provided (for usage across several indices) or created internally. The caching dir will also create a wrapped IndexOutput, that will flush to the main dir if the allocator can no longer provide byte[] (exhausted). When a file is flushed from the cache to the main directory, it will return all the currently allocated byte[] to the BufferAllocator to be reused by other files. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-3425) NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control
[ https://issues.apache.org/jira/browse/LUCENE-3425?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13765336#comment-13765336 ] Michael McCandless commented on LUCENE-3425: It's true that open SegmentReaders will hold onto those RAMFiles. However, those segments should be the smallish ones and they should be merged away. It's odd that in your case you see them lasting so long. Can you give more details about how you're using it? Have you changed the default MergePolicy/settings? Are you closing old readers after reopening new ones? NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control Key: LUCENE-3425 URL: https://issues.apache.org/jira/browse/LUCENE-3425 Project: Lucene - Core Issue Type: Improvement Components: core/index Affects Versions: 3.4, 4.0-ALPHA Reporter: Shay Banon Fix For: 5.0, 4.5 A discussion on IRC raised several improvements that can be made to NRT caching dir. Some of the problems it currently has are: 1. Not explicitly controlling the memory usage, which can result in overusing memory (for example, large new segments being committed because refreshing is too far behind). 2. Heap fragmentation because of constant allocation of (probably promoted to old gen) byte buffers. 3. Not being able to control the memory usage across indices for multi index usage within a single JVM. A suggested solution (which still needs to be ironed out) is to have a BufferAllocator that controls allocation of byte[], and allow to return unused byte[] to it. It will have a cap on the size of memory it allows to be allocated. The NRT caching dir will use the allocator, which can either be provided (for usage across several indices) or created internally. The caching dir will also create a wrapped IndexOutput, that will flush to the main dir if the allocator can no longer provide byte[] (exhausted). When a file is flushed from the cache to the main directory, it will return all the currently allocated byte[] to the BufferAllocator to be reused by other files. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-3425) NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control
[ https://issues.apache.org/jira/browse/LUCENE-3425?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13103723#comment-13103723 ] Michael McCandless commented on LUCENE-3425: Actually, NRTCachingDir does explicitly control the RAM usage in that if its cache is using too much RAM then the next createOutput will go straight to disk. The one thing it does not do is evict the created files after they close. So, if you flush a big segment in IW, then NRTCachingDir will keep those files in RAM even though its now over-budget. (But the next segment to flush will go straight to disk). I think this isn't that big a problem in practice; ie, as long as you set your IW RAM buffer to something not too large, or you ensure you are opening a new NRT reader often enough that the accumulated docs won't create a very large segment, then the excess RAM used by NRTCachingDir will be bounded. Still it would be nice to fix it so it evicts the files that set it over, such that it's always below the budget once the outputs is closed. And I agree we should make it possible to have a single pool for accounting purposes, so you can share this pool across multiple NRTCachingDirs (and other things that use RAM). NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control Key: LUCENE-3425 URL: https://issues.apache.org/jira/browse/LUCENE-3425 Project: Lucene - Java Issue Type: Improvement Components: core/index Reporter: Shay Banon A discussion on IRC raised several improvements that can be made to NRT caching dir. Some of the problems it currently has are: 1. Not explicitly controlling the memory usage, which can result in overusing memory (for example, large new segments being committed because refreshing is too far behind). 2. Heap fragmentation because of constant allocation of (probably promoted to old gen) byte buffers. 3. Not being able to control the memory usage across indices for multi index usage within a single JVM. A suggested solution (which still needs to be ironed out) is to have a BufferAllocator that controls allocation of byte[], and allow to return unused byte[] to it. It will have a cap on the size of memory it allows to be allocated. The NRT caching dir will use the allocator, which can either be provided (for usage across several indices) or created internally. The caching dir will also create a wrapped IndexOutput, that will flush to the main dir if the allocator can no longer provide byte[] (exhausted). When a file is flushed from the cache to the main directory, it will return all the currently allocated byte[] to the BufferAllocator to be reused by other files. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-3425) NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control
[ https://issues.apache.org/jira/browse/LUCENE-3425?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13101567#comment-13101567 ] Michael McCandless commented on LUCENE-3425: Also, a quick win on trunk is to use IOCtx's FlushInfo.estimatedSegmentSize to decide up front whether to try caching or not. Ie if the to-be-flushed segment is too large we should not cache it. NRT Caching Dir to allow for exact memory usage, better buffer allocation and global cross indices control Key: LUCENE-3425 URL: https://issues.apache.org/jira/browse/LUCENE-3425 Project: Lucene - Java Issue Type: Improvement Components: core/index Reporter: Shay Banon A discussion on IRC raised several improvements that can be made to NRT caching dir. Some of the problems it currently has are: 1. Not explicitly controlling the memory usage, which can result in overusing memory (for example, large new segments being committed because refreshing is too far behind). 2. Heap fragmentation because of constant allocation of (probably promoted to old gen) byte buffers. 3. Not being able to control the memory usage across indices for multi index usage within a single JVM. A suggested solution (which still needs to be ironed out) is to have a BufferAllocator that controls allocation of byte[], and allow to return unused byte[] to it. It will have a cap on the size of memory it allows to be allocated. The NRT caching dir will use the allocator, which can either be provided (for usage across several indices) or created internally. The caching dir will also create a wrapped IndexOutput, that will flush to the main dir if the allocator can no longer provide byte[] (exhausted). When a file is flushed from the cache to the main directory, it will return all the currently allocated byte[] to the BufferAllocator to be reused by other files. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org