com.bigdata.rdf.internal.HashCollisionUtility Maven / Gradle / Ivy
Show all versions of bigdata-rdf-test Show documentation
package com.bigdata.rdf.internal;
import java.io.BufferedInputStream;
import java.io.File;
import java.io.FileInputStream;
import java.io.IOException;
import java.io.InputStream;
import java.lang.reflect.InvocationTargetException;
import java.nio.ByteBuffer;
import java.security.NoSuchAlgorithmException;
import java.util.Arrays;
import java.util.LinkedHashMap;
import java.util.LinkedHashSet;
import java.util.LinkedList;
import java.util.List;
import java.util.Map;
import java.util.Properties;
import java.util.Set;
import java.util.TimeZone;
import java.util.UUID;
import java.util.concurrent.BlockingQueue;
import java.util.concurrent.Callable;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.FutureTask;
import java.util.concurrent.LinkedBlockingQueue;
import java.util.concurrent.RejectedExecutionException;
import java.util.concurrent.ThreadPoolExecutor;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicBoolean;
import java.util.concurrent.atomic.AtomicLong;
import java.util.concurrent.locks.ReentrantLock;
import java.util.zip.Deflater;
import java.util.zip.GZIPInputStream;
import org.apache.log4j.Logger;
import org.openrdf.model.BNode;
import org.openrdf.model.Literal;
import org.openrdf.model.Statement;
import org.openrdf.model.URI;
import org.openrdf.model.Value;
import org.openrdf.rio.RDFFormat;
import org.openrdf.rio.RDFHandlerException;
import org.openrdf.rio.RDFParseException;
import org.openrdf.rio.RDFParser;
import org.openrdf.rio.RDFParserFactory;
import org.openrdf.rio.RDFParserRegistry;
import org.openrdf.rio.helpers.RDFHandlerBase;
import com.bigdata.Banner;
import com.bigdata.btree.BTree;
import com.bigdata.btree.DefaultTupleSerializer;
import com.bigdata.btree.IRangeQuery;
import com.bigdata.btree.ITuple;
import com.bigdata.btree.ITupleIterator;
import com.bigdata.btree.IndexMetadata;
import com.bigdata.btree.keys.DefaultKeyBuilderFactory;
import com.bigdata.btree.keys.IKeyBuilder;
import com.bigdata.btree.keys.KV;
import com.bigdata.btree.keys.KeyBuilder;
import com.bigdata.btree.keys.SuccessorUtil;
import com.bigdata.btree.raba.codec.CanonicalHuffmanRabaCoder;
import com.bigdata.btree.raba.codec.FrontCodedRabaCoder;
import com.bigdata.io.ByteArrayBuffer;
import com.bigdata.io.DataOutputBuffer;
import com.bigdata.io.DirectBufferPool;
import com.bigdata.io.compression.RecordCompressor;
import com.bigdata.journal.BufferMode;
import com.bigdata.journal.Journal;
import com.bigdata.rdf.internal.impl.BlobIV;
import com.bigdata.rdf.internal.impl.literal.PartlyInlineTypedLiteralIV;
import com.bigdata.rdf.internal.impl.uri.PartlyInlineURIIV;
import com.bigdata.rdf.model.BigdataValue;
import com.bigdata.rdf.model.BigdataValueFactory;
import com.bigdata.rdf.model.BigdataValueFactoryImpl;
import com.bigdata.rdf.model.BigdataValueSerializer;
import com.bigdata.rdf.store.AbstractTripleStore;
import com.bigdata.rdf.vocab.BaseVocabulary;
import com.bigdata.rdf.vocab.DefaultBigdataVocabulary;
import com.bigdata.rwstore.sector.IMemoryManager;
import com.bigdata.rwstore.sector.MemoryManager;
import com.bigdata.util.Bytes;
import com.bigdata.util.BytesUtil;
import com.bigdata.util.concurrent.Latch;
/**
* Utility class to parse some RDF resource(s) and count hash collisions using a
* variety of hash codes.
*
* TODO Various data sets:
*
*
* /nas/data/lubm/U1/data/University0/
* /nas/data/bsbm/bsbm_2785/dataset.nt.gz
* /nas/data/bsbm/bsbm_566496/dataset.nt.gz
* /data/bsbm3_200m_1MSplits
*
* 8B triple bioinformatics data set.
*
* BTC data (some very large literals)
*
*
*
* TODO order preserving hash codes could be interesting here. Look at 32 and 64
* bit variants of the math and at generalized order preserving hash codes. With
* order preserving hash codes, it makes sense to insert all Unicode terms into
* TERM2ID such that we have a total order there.
*
* TODO benchmark the load time with different hash codes. the cost of the hash
* computation and the randomness of the distribution will both play a role. The
* B+Tree will need to be setup with a sufficient [writeRetentionQueue] and we
* will need to specify [-server -Xmx1G].
*
* SHA-256 - no collisions on BSBM 200M. 30G file. time?
*
* 32-bit hash codes. #collisions=1544132 Elapsed: 16656445ms Journal size:
* 23841341440 bytes (23G)
*
* Now limiting the size of values in a leaf and also increasing the branching
* factor to 512 (was 32). [The current run is scanning after the initial
* insert, which involves a little wasted effort. It was also without the
* -server -Xmx2g, and write retention queue parameters. Finally, it was
* serializing BigdataValue objects, including their IV, rather than RDF Value
* objects. The code has since been modified to serialize just the BigdataValue
* Also, I've since raised the initial extent from 10M to 200M].
* maxCollisions=3, Elapsed: 22579073ms Journal size: 35270950912 bytes
*
* Now buffering 100k values at a time: 2x faster.
*
*
* U1:
* Elapsed: 23379ms
* NumStatements: 1000313
* NumDistinctVals: 291259
* TotalKeyBytes: 1747554
* TotalValBytes: 60824514
* MaxCollisions: 1
* TotalCollisions: 6
* Journal size: 209715200 bytes
* name m height nnodes nleaves nodeBytes leafBytes totalBytes avgNodeBytes avgLeafBytes minNodeBytes maxNodeBytes minLeafBytes maxLeafBytes
* lex 1024 1 1 474 7913 3662623 3670536 7913 7727 7913 7913 5786 13784
*
*
* With only a byte (versus short) counter in the key. Oddly, this has no impact
* on the average leaf size. That suggests that the keys in the leaves are very
* sparse in terms of the hash code space such that prefix compression is not
* really doing that much for us.
*
*
* Elapsed: 23235ms
* NumStatements: 1000313
* NumDistinctVals: 291259
* TotalKeyBytes: 1456295
* TotalValBytes: 60824514
* MaxCollisions: 1
* TotalCollisions: 6
* Journal size: 209715200 bytes
* name m height nnodes nleaves nodeBytes leafBytes totalBytes avgNodeBytes avgLeafBytes minNodeBytes maxNodeBytes minLeafBytes maxLeafBytes
* lex 1024 1 1 474 7913 3371370 3379283 7913 7112 7913 7913 5274 12774
*
*
* BSBM 200M: This is the best time and space so far. using a byte counter
* rather than a short.
*
*
* Elapsed: 16338357ms
* NumStatements: 198808848
* NumDistinctVals: 45647082
* TotalKeyBytes: 228235410
* TotalValBytes: 11292849582
* MaxCollisions: 3
* TotalCollisions: 244042
* Journal size: 16591683584 bytes
*
*
* BSBM 200M: Note: I restarted this run after terminating yourkit so the
* results should be valid (right?). The main changes are to use stringValue()
* to test for dateTime, to use the canonical huffman coder for the leaf keys.
*
*
* Elapsed: 20148506ms
* NumStatements: 198808848
* NumDistinctVals: 45647082
* TotalKeyBytes: 228235410
* TotalValBytes: 11292849582
* MaxCollisions: 3
* TotalCollisions: 244042
* Journal size: 16591683584 bytes
*
*
* BSBM 200M: raw records are compress if they are over 64 bytes long.
*
*
* Elapsed: 18757003ms
* NumStatements: 198808848
* NumDistinctVals: 45647082
* TotalKeyBytes: 228235410
* TotalValBytes: 7910596818
* MaxCollisions: 3
* TotalCollisions: 244042
* Journal size: 12270108672 bytes
*
*
* BSBM 200M: literals LT 64 byte labels are assumed inlined into statement
* indices (except datatype URIs).
*
*
* Elapsed: 16193915ms
* NumStatements: 198808848
* NumDistinctVals: 43273381
* NumShortLiterals: 2723662
* TotalKeyBytes: 216366905
* TotalValBytes: 7807037644
* MaxCollisions: 3
* TotalCollisions: 219542
* Journal size: 11083186176 bytes
*
*
* BSBM 200M: uris LT 64 byte localNames are assumed inlined into statement
* indices (plus datatype literals LT 64 bytes).
*
*
* Elapsed: 5699248ms
* NumStatements: 198808848
* NumDistinctVals: 12198222
* NumShortLiterals: 32779032
* NumShortURIs: 493520581
* TotalKeyBytes: 60991110
* TotalValBytes: 4944223808
* MaxCollisions: 2
* TotalCollisions: 17264
* Journal size: 7320764416 bytes
*
*
* BSBM 200M: one parser thread and one indexer thread.
*
*
* Elapsed: 3724415ms
* NumStatements: 198808848
* NumDistinctVals: 12198222
* NumShortLiterals: 32779032
* NumShortBNodes: 0
* NumShortURIs: 493520581
* TotalKeyBytes: 60991110
* TotalValBytes: 4944223808
* MaxCollisions: 2
* TotalCollisions: 17264
* Journal size: 7320764416 bytes
*
*
* GC OH problem trying to run multiple parsers against BSBM 200M when split
* into 200 files.
*
*
* valBufSize := 10000
* valQueueCapacity = 100
* maxDrain := 50
* nparserThreads := 2
* parserWorkQueue := 1000
*
*
* BSBM 200M - this is 3x longer. This run did not have the GC OH problem, but
* GC had frequent 10% spikes, which is a lot in comparison to our best run.
*
*
* valBufSize := 1000
* valQueueCapacity = 10
* maxDrain := 5
* nparserThreads := 4
* parserWorkQueue := 100
*
* Elapsed: 9589520ms
* NumStatements: 198805837
* NumDistinctVals: 12202052
* NumShortLiterals: 32776100
* NumShortBNodes: 0
* NumShortURIs: 493514954
* TotalKeyBytes: 61010260
* TotalValBytes: 4945278396
* MaxCollisions: 2
* TotalCollisions: 17260
* Journal size: 7320764416 bytes
*
*
* BSBM 200M: split in 200 files. 69m versus best time so far of 62m. There is
* only one thread in the pool, but the caller runs policy means that we are
* actually running two parsers. So, this is not really the same as the best
* run, which was one parser running in the main thread with the indexer running
* in another thread.
*
*
* valBufSize := 10000
* valQueueCapacity = 10
* maxDrain := 5
* nparserThreads := 1
* parserWorkQueue := 100
*
* Elapsed: 4119775ms
* NumStatements: 198805837
* NumDistinctVals: 12202052
* NumShortLiterals: 32776100
* NumShortBNodes: 0
* NumShortURIs: 493514954
* TotalKeyBytes: 61010260
* TotalValBytes: 4945278396
* MaxCollisions: 2
* TotalCollisions: 17260
* Journal size: 7320764416 bytes
*
*
* BSBM 200M with 1M statement splits using the memory manager to buffer the
* data on the native heap. This is the best score so far (compare with
* 3724415ms with one parser and one indexer thread). For some reason, the #of
* distinct values and literals is slightly different for these two runs. One
* other change in this run is that we always gzip the record since we can not
* deserialize the record unless we know in advance whether or not it is
* compressed. Previous runs had conditionally compressed based on the original
* byte[] value length and stored the compressed record iff it was shorter.
* However, we can only conditionally compress if we use a header or bit flag to
* indicate that the record is compressed. Peak memory manager use was 262M.
*
*
* Elapsed: 2863898ms
* NumStatements: 198805837
* NumDistinctVals: 12,202,052
* NumShortLiterals: 61,100,900
* NumShortBNodes: 0
* NumShortURIs: 493514954
* TotalKeyBytes: 61010260
* TotalValBytes: 4945376779
* MaxCollisions: 2
* TotalCollisions: 17260
* Journal size: 7320764416 bytes
*
*
* BSBM 200M using memory manager (high tide of 351M) and 5 parser threads (plus
* the main thread). Heap usage is pretty controlled.
*
*
* Elapsed: 2803451ms
* NumStatements: 198805837
* NumDistinctVals: 12202052
* NumShortLiterals: 61100900
* NumShortBNodes: 0
* NumShortURIs: 493514954
* TotalKeyBytes: 61010260
* TotalValBytes: 4945376779
* MaxCollisions: 2
* TotalCollisions: 17260
* Journal size: 7320764416 bytes
*
*
* BSBM 200M. Using memory manager and only one parser thread. This does run
* significantly slower (55m versus 47m with two parser threads). It might not
* be slower if we also ran against the single source file (this ran against the
* split files) since each chunk placed onto the queue would then be full, but I
* doubt that this will make that much difference.
*
*
* Elapsed: 3300871ms
* NumStatements: 198805837
* NumDistinctVals: 12049125
* NumShortLiterals: 61100900
* NumShortBNodes: 0
* NumShortURIs: 493514954
* TotalKeyBytes: 60245625
* TotalValBytes: 4877760110
* MaxCollisions: 2
* TotalCollisions: 16840
* Journal size: 7320764416 bytes
*
*
* BSBM 200M. Using memory manager, one parser thread (the caller), and a single
* source file. The question is whether we do better with a statement handler
* that is only flushed incrementally (when full) compared to using 2 parsers
* and flushing each time we reach the end of a 1M statement source file. Nope.
* This was 77 minutes. (This was a fair comparison since the source files for
* the split sources are compressed. So we really do better with two parsers and
* split files)
*
*
* /allocationCount=0
* /bufferCapacity=1000
* /bufferCount=232
* /extent=243269632
* /slotBytes=0
* /userBytes=0
* Elapsed: 4605950ms
* NumStatements: 198808848
* NumDistinctVals: 12198222
* NumShortLiterals: 61103832
* NumShortBNodes: 0
* NumShortURIs: 493520581
* TotalKeyBytes: 60991110
* TotalValBytes: 4944322031
* MaxCollisions: 2
* TotalCollisions: 17264
* Journal size: 7320764416 bytes
*
*
* TODO Try with only N bytes worth of the SHA hash code, leaving some bits left
* over for partitioning URIs, Literals, and BNodes (for told bnode mode) and
* for a counter to break ties when there is a hash collision. We should wind up
* with an 8-12 byte termId which is collision proof and very well distributed.
*
* TODO Add bit flags at the front for {BLOB, URI, Literal, BNode} (BLOB being
* the odd one out). If we move BLOBs out of the key range of other plain
* literals, or literals of a given language code or datatype, then we can not
* do an ordered scan of the literals anymore which is inclusive of the blobs.
* There is a similar consequence of moving small literals into the statement
* index.
*
* If we inline small unicode values (<32 bytes) and reserve the TERM2ID index
* for large(r) values then we can approach a situation in which it serves
* solely for blobs but with a tradeoff in size (of the statement indices)
* versus indirection.
*
* Large value promotion does not really let us handle large blobs
* (multi-megabytes) in s/o as a 50 4M blobs would fill up a shard. There, I
* think that we need to give the control over to the application and require it
* to write on a shared resource (shared file system, S3, etc). The value
* inserted into the index would then be just the pathname in the shared file
* system or the URL of the S3 resource. This breaks the ACID decision boundary
* though as the application has no means available to atomically decide that
* the resource does not exist and hence create it. Even using a conditional
* E-Tag on S3 would not work since it would have to have an index over the S3
* entities to detect a write-write conflict for the same data under different
* URLs.
*
* @author thompsonbry
*/
public class HashCollisionUtility {
private final static Logger log = Logger
.getLogger(HashCollisionUtility.class);
/**
* An index mapping hashCode(Value)+counter : Value. This
* provides a dictionary for RDF {@link Value}s encountered when loading
* {@link Statement}s into the database. The counter provides a simple
* mechanism for reconciling hash collisions.
*/
private final BTree termsIndex;
private final LexiconConfiguration conf;
private final BigdataValueFactory vf;
/**
* Counters for things that we track.
*/
private static class Counters {
/**
* #of statements visited.
*/
private final AtomicLong nstmts = new AtomicLong();
/**
* The #of {@link URI}s whose localName was short enough
* that we decided to inline them into the statement indices instead.
*/
private final AtomicLong nshortURIs = new AtomicLong();
/**
* The #of {@link BNode}s whose ID was short enough that we
* decided to inline them into the statement indices instead (this also
* counts blank nodes which are inlined because they have integer or
* UUID IDs).
*/
private final AtomicLong nshortBNodes = new AtomicLong();
/**
* The #of {@link Literal}s which were short enough that we decided to
* inline them into the statement indices instead.
*/
private final AtomicLong nshortLiterals = new AtomicLong();
// private final ConcurrentWeakValueCacheWithBatchedUpdates valueCache;
/**
* The size of the hash collision set for the RDF Value with the most
* hash collisions observed to date.
*/
private final AtomicLong maxCollisions = new AtomicLong();
/**
* The total #of hash collisions.
*/
private final AtomicLong totalCollisions = new AtomicLong();
// /**
// * The #of RDF {@link Value}s which were found in the {@link
// #valueCache},
// * thereby avoiding a lookup against the index.
// */
// private final AtomicLong ncached = new AtomicLong();
/**
* The #of distinct RDF {@link Value}s inserted into the index.
*/
private final AtomicLong ninserted = new AtomicLong();
/** The total #of bytes in the generated B+Tree keys (leaves only). */
private final AtomicLong totalKeyBytes = new AtomicLong();
/** The total #of bytes in the serialized RDF Values. */
private final AtomicLong totalValBytes = new AtomicLong();
} // class Counters
//// private interface IHashCode {
//// void hashCode(IKeyBuilder keyBuilder,Object o);
//// }
//
// private static class Int32HashCode { //implements IHashCode {
//
// public void hashCode(IKeyBuilder keyBuilder, Object o) {
//
// keyBuilder.append(o.hashCode());
//
// }
//
// }
//
// private static class MessageDigestHashCode { //implements IHashCode {
//
// final MessageDigest d;
//
// MessageDigestHashCode() throws NoSuchAlgorithmException {
//
// d = MessageDigest.getInstance("SHA-256"); // 256 bits (32 bytes)
//
// }
//
// public void hashCode(IKeyBuilder keyBuilder, final byte[] b) {
//
// d.reset();
// d.digest(b);
// keyBuilder.append(d.digest());
//
// }
//
// }
/**
* Lock used to coordinate {@link #shutdown()} and the {@link #valueQueue}.
*/
private final ReentrantLock lock = new ReentrantLock();
/**
* Latch which is incremented as we accept files to parse and decremented
* once a parser begins to parse that file.
*/
private final Latch parserQueueLatch = new Latch(lock);
/**
* Latch which is incremented once we begin to parse a file and decremented
* as the parser task completes.
*/
private final Latch parserRunLatch = new Latch(lock);
/**
* Thread pool used to run the parser.
*/
private final ExecutorService parserService;
/**
* Thread pool used to run the parser and indexer.
*/
private final ExecutorService indexerService;
/**
* Class hooks the runnable to provide reporting on the outcome of the
* {@link FutureTask}.
*/
private class ReportingFutureTask extends FutureTask {
public final File file;
public ReportingFutureTask(final File file, Callable callable) {
super(callable);
this.file = file;
parserQueueLatch.inc();
}
public void run() {
try {
parserRunLatch.inc();
parserQueueLatch.dec();
super.run();
parserRunLatch.dec();
} finally {
report(this);
}
}
/**
* Callback is invoked when a {@link ParseFileTask} completes.
*
* @param task
* The future for that task.
*/
protected void report(final ReportingFutureTask task) {
try {
task.get();
if (log.isDebugEnabled())
log.debug("Finished parsing: " + task.file
+ ", queueLatch=" + parserQueueLatch
+ ", runLatch=" + parserRunLatch);
} catch (ExecutionException ex) {
log.error(ex, ex);
} catch (InterruptedException e) {
// propagate the interrupt.
Thread.currentThread().interrupt();
}
}
}
/**
* A {@link Bucket} has an unsigned byte[] key and an unordered
* list of long addrs for byte[] values.
* {@link Bucket} implements {@link Comparable} can can be used to place an
* array of {@link Bucket}s into ascending key order.
*
* TODO This is space efficient for large {@link Value}s, but it would not
* be efficient for storing binding sets which hash to the same key. In the
* case of binding sets, the binding sets are normally small. An extensible
* hash table would conserve space by dynamically determining the #of hash
* bits in the address, and hence mapping the records onto a smaller #of
* pages.
*/
static private class Bucket implements Comparable {
/**
* The unsigned byte[] key.
*/
public final byte[] key;
/**
* The list of addresses for this bucket.
*
* TODO Collisions in a bucket are very rare given an int32 hash code,
* so this should be optimized for the common case with a single
* address.
*/
public final List addrs = new LinkedList();
public Bucket(final byte[] key) {
if(key == null)
throw new IllegalArgumentException();
this.key = key;
}
public Bucket(final byte[] key,final long addr) {
this(key);
addrs.add(addr);
}
/**
* Add an address to this bucket.
*
* @param addr
* The address.
*/
public void add(final long addr) {
addrs.add(addr);
}
/**
* Order {@link Bucket}s into ascending unsigned byte[] key
* order.
*/
public int compareTo(final Bucket o) {
return BytesUtil.compareBytes(key, o.key);
}
}
/**
* A chunk of RDF {@link Value}s from the parser which are ready to be
* inserted into the TERMS index.
*/
static private class ValueBuffer {
/**
* The allocation contexts which can be released once these data have
* been processed.
*/
private final Set contexts = new LinkedHashSet();
/**
* The #of distinct records in the addrMap (this is more than the map
* size if there are hash collisions since some buckets will have more
* than one entry).
*/
private final int nvalues;
/**
* A map from the unsigned byte[] keys to the collision
* bucket containing the address of each record for a given
* unsigned byte[] key.
*/
private final Map addrMap;
/**
*
* @param contexts
* The allocation contexts for the records in the addrMap.
* @param nvalues
* The #of distinct records in the addrMap (this is more than
* the map size if there are hash collisions since some
* buckets will have more than one entry).
* @param addrMap
* A map from the unsigned byte[] keys to the
* collision bucket containing the address of each record for
* a given unsigned byte[] key.
*/
public ValueBuffer(final List contexts,
final int nvalues, final Map addrMap) {
if (contexts == null)
throw new IllegalArgumentException();
if (addrMap == null)
throw new IllegalArgumentException();
this.contexts.addAll(contexts);
this.nvalues = nvalues;
this.addrMap = addrMap;
}
/**
* Clear the address map and the {@link IMemoryManager} allocation
* contexts against which the data were stored.
*/
public void clear() {
addrMap.clear();
for(IMemoryManager context : contexts) {
context.clear();
}
}
public long getUserBytes() {
long nbytes = 0L;
for(IMemoryManager context : contexts) {
nbytes += context.getUserBytes();
}
return nbytes;
}
} // class ValueBuffer
/**
* Queue used to hand off {@link ValueBuffer}s from the parser to the
* indexer.
*/
private BlockingQueue valueQueue;
/**
* Counters for things that we track.
*/
private final Counters c = new Counters();
/**
* The upper bound on the size of a {@link ValueBuffer} chunk (currently in
* slotBytes for the allocations against the {@link MemoryManager}).
*
* The size of the chunks, the capacity of the queue, and the number of
* chunks that may be combined into a single chunk for the indexer may all
* be used to adjust the parallelism and efficiency of the parsing and
* indexing. You have to be careful not to let too much data onto the Java
* heap, but the indexer will do better when it is given a bigger chunk
* since it can order the data and be more efficient in the index updates.
*/
final int valBufSize = Bytes.megabyte32 * 10;// 100000;
/** Capacity of the {@link #valueQueue}. */
final int valQueueCapacity = 10;
/**
* Maximum #of chunks to drain from the {@link #valueQueue} in one go. This
* bounds the largest chunk that we will index at one go. You can remove the
* limit by specifying {@link Integer#MAX_VALUE}.
*/
final int maxDrain = 5;
/**
* The size of the read buffer when reading a file.
*/
final int fileBufSize = 1024 * 8;// default 8k
/**
* How many parser threads to use. There can be only one parser per file,
* but you can parse more than one file at a time.
*/
final int nparserThreads = 1;
/**
* The size of the work queue for the {@link #parserService}.
*
* Note: This should be large enough that we will not wait around forever if
* the caller is forced to parse a file rather than scan the file system for
* the next file to be parsed. This hack is introduced by the need to handle
* a {@link RejectedExecutionException} from the {@link #parserService}. We
* do that by forcing the parse task to run in the caller's thread. Another
* choice would be for the caller to catch the
* {@link RejectedExecutionException}, wait a bit, and then retry.
*/
final int parserWorkQueueCapacity = 100;
/**
* A direct memory heap used to buffer RDF {@link Value}s which will be
* inserted into the TERMS index. A distinct child {@link IMemoryManager}
* context is created by the {@link StatementHandler} each time it needs to
* buffer data. The {@link StatementHandler} monitors the size of the
* allocation context to decide when it is "big enough" to be transferred
* onto the {@link #valueQueue}. The indexer eventually obtains the context
* from the {@link #valueQueue}. Once the indexer is done with a context, it
* {@link IMemoryManager#clear() clears} the context. The total memory
* across the allocation contexts is released back to the
* {@link DirectBufferPool} in {@link #shutdown()} and
* {@link #shutdownNow()} and no later than when the {@link #mmgr} is
* finalized.
*/
final MemoryManager mmgr;
/**
* The #of buffers to give to the {@link MemoryManager}.
*/
private final int nbuffers = 1000;
private HashCollisionUtility(final Journal jnl) {
this.termsIndex = getTermsIndex(jnl);
/*
* Setup the parser thread pool. If there is an attempt to run more
* threads then
*
* Note: The pool size is one less than the total #of specified threads
* since the caller will wind up running tasks rejected by the pool. If
* the pool would be empty then it is [null] and the caller will run
* the parser in its own thread.
*
* Note: The work queue is bounded so that we do not read any too far in
* the file system. The #of threads is bounded so that we do not run too
* many parsers at once. However, running multiple parsers can increase
* throughput as the parser itself caps out at ~ 68k tps.
*/
if (nparserThreads > 1) {
// this.parserService =
// Executors.newFixedThreadPool(nparserThreads);
final int corePoolSize = nparserThreads-1;
final int maximumPoolSize = nparserThreads-1;
final long keepAliveTime = 60;
final TimeUnit unit = TimeUnit.SECONDS;
final BlockingQueue workQueue = new LinkedBlockingQueue(
parserWorkQueueCapacity);
// final BlockingQueue workQueue = new SynchronousQueue();
this.parserService = new ThreadPoolExecutor(corePoolSize,
maximumPoolSize, keepAliveTime, unit, workQueue,
new ThreadPoolExecutor.CallerRunsPolicy()
);
} else {
/*
* The caller must run the parser in its own thread.
*/
this.parserService = null;
}
// But they all feed the same indexer.
this.indexerService = Executors.newSingleThreadExecutor();
// *blocking* queue of ValueBuffers to be indexed
this.valueQueue = new LinkedBlockingQueue(
valQueueCapacity);// lock);
vf = BigdataValueFactoryImpl.getInstance("test");
final BaseVocabulary vocab;
try {
vocab = (BaseVocabulary) Class.forName(
AbstractTripleStore.Options.DEFAULT_VOCABULARY_CLASS)
.getDeclaredConstructor(String.class)
.newInstance(vf.getNamespace());
vocab.init();
} catch (Exception e) {
throw new RuntimeException(e);
}
// factory does not support any extensions.
final IExtensionFactory xFactory = new IExtensionFactory() {
@Override
public void init(final IDatatypeURIResolver resolver,
final ILexiconConfiguration config) {
// NOP
}
@Override
@SuppressWarnings("rawtypes")
public IExtension[] getExtensions() {
return new IExtension[] {};
}
};
final InlineURIFactory uriFactory = new InlineURIFactory();
uriFactory.init(vocab);
/*
* Note: This inlines everything *except* xsd:dateTime, which
* substantially reduces the data we will put into the index.
*
* @todo Do a special IExtension implementation to handle xsd:dateTime
* since the DateTimeExtension uses the LexiconRelation to do its work.
*/
conf = new LexiconConfiguration(
256, // blobsThreshold
true, // inlineXSDDatatypeLiterals
true, // inlineTextLiterals
64, // maxInlineStringLength
true, // inlineBNodes
false, // inlineDateTimes
TimeZone.getDefault(), // inlineDateTimesTimeZone
false, // rejectInvalidXSDValues
xFactory, // extension factory
vocab, // predefined vocabulary
vf,
uriFactory,
false, // GeoSpatial support
null // GeoSpatial config string
);
// valueCache = new ConcurrentWeakValueCacheWithBatchedUpdates(
// 50000 // hard reference queue capacity
// );
mmgr = new MemoryManager(DirectBufferPool.INSTANCE, nbuffers);
}
/**
* Start the task which will index data as it is parsed.
*/
public void start() {
lock.lock();
try {
if (indexerTask != null)
throw new IllegalStateException();
// start indexer.
indexerTask = new FutureTask(new IndexerMainTask());
indexerService.submit(indexerTask);
// allow parsers to run.
parsing.set(true);
} finally {
lock.unlock();
}
}
/**
* Future for the task which drains the {@link #valueQueue} and indexes
* the {@link ValueBuffer}s drained from that queue.
*/
private FutureTask indexerTask;
/** Flag is true while parsers are still running. */
private final AtomicBoolean parsing = new AtomicBoolean(false);
/**
* Poison pill used to indicate that no more objects will be placed onto the
* {@link #valueQueue}.
*/
private final ValueBuffer poisonPill = new ValueBuffer(
new LinkedList(), 0,
new LinkedHashMap());
/**
* Normal shutdown. Running parsers will complete and their data will be
* indexed, but new parsers will not start. This method will block until
* all data has been indexed.
*
* @throws Exception
*/
public void shutdown() throws Exception {
log.debug("shutting down...");
lock.lock();
try {
if (log.isDebugEnabled())
log.debug("Waiting on parserQueueLatch: " + parserQueueLatch);
parserQueueLatch.await();
if (parserService != null) {
// no new parsers may start
parserService.shutdown();
}
if (log.isDebugEnabled())
log.debug("Waiting on parserRunLatch: " + parserRunLatch);
parserRunLatch.await();
// no parsers should be running.
parsing.set(false);
// drop a poison pill on the queue.
log.debug("Inserting poison pill.");
valueQueue.put(poisonPill);
if (indexerTask != null) {
// wait for the indexer to finish.
indexerTask.get();
}
if (indexerService != null)
indexerService.shutdown();
if (mmgr != null) {
if (log.isInfoEnabled())
log.info(mmgr.getCounters().toString());
mmgr.clear();
}
} finally {
lock.unlock();
}
log.debug("all done.");
}
/**
* Immediate shutdown. Running tasks will be canceled.
*
* @throws Exception
*/
public void shutdownNow() throws Exception {
log.debug("shutdownNow");
parsing.set(false);
if (parserService != null)
parserService.shutdownNow();
if (indexerService != null)
indexerService.shutdownNow();
if (indexerTask != null) {
indexerTask.cancel(true/* mayInterruptIfRunning */);
}
if (mmgr != null) {
mmgr.clear();
}
}
/**
* Task drains the valueQueue and runs an {@link IndexerTask} each time
* something is drained from that queue.
*
* @author thompsonbry
*/
private class IndexerMainTask implements Callable {
public Void call() throws Exception {
boolean done = false;
while (!done) {
try {
// Blocking take so we know that there is something ready.
final ValueBuffer first = valueQueue.take();
// Drain queue, but keep an eye out for that poison pill.
final LinkedList coll = new LinkedList();
// The element we already took from the queue.
coll.add(first);
// Drain (non-blocking).
final int ndrained = valueQueue.drainTo(coll, maxDrain) + 1;
if (log.isInfoEnabled())
log.info("Drained " + ndrained + " chunks with "
+ valueQueue.size()
+ " remaining in the queue.");
// look for and remove poison pill, noting if found.
if (coll.remove(poisonPill)) {
if (log.isDebugEnabled())
log.debug("Found poison pill.");
done = true;
// fall through and index what we already have.
}
if (!coll.isEmpty()) {
// combine the buffers into a single chunk.
final ValueBuffer b = combineChunks(coll);
if (log.isDebugEnabled())
log.debug("Will index " + coll.size()
+ " chunks having " + b.nvalues
+ " values in " + b.getUserBytes()
+ " bytes");
// Now index that chunk.
new IndexValueBufferTask(mmgr, b, termsIndex, vf, c)
.call();
}
} catch (Throwable t) {
log.error(t, t);
HashCollisionUtility.this.shutdownNow();
throw new RuntimeException(t);
}
} // while(!done)
log.debug("done.");
return (Void) null;
}
/**
* Combine chunks from the queue into a single chunk.
*/
private ValueBuffer combineChunks(final LinkedList coll) {
final ValueBuffer b;
if (coll.size() == 1) {
// There is only one chunk.
b = coll.getFirst();
} else {
// Combine together into a single chunk.
int nvalues = 0;
for (ValueBuffer t : coll)
nvalues += t.nvalues;
final List contexts = new LinkedList();
final LinkedHashMap addrMap = new LinkedHashMap();
// int off = 0;
for (ValueBuffer t : coll) {
contexts.addAll(t.contexts);
nvalues += t.nvalues;
for(Bucket bucket : t.addrMap.values()) {
final Bucket tmp = addrMap.get(bucket.key);
if(tmp == null) {
// copy bucket.
addrMap.put(bucket.key, bucket);
} else {
// merge bucket.
tmp.addrs.addAll(bucket.addrs);
}
}
// System
// .arraycopy(t.keys/* src */, 0/* srcPos */,
// keys/* dest */, off/* destPos */,
// t.nvalues/* length */);
//
// System
// .arraycopy(t.addrs/* src */, 0/* srcPos */,
// addrs/* dest */, off/* destPos */,
// t.nvalues/* length */);
//
// off += t.nvalues;
}
b = new ValueBuffer(contexts, nvalues, addrMap);
}
return b;
}
} // class IndexerMainTask
/**
* Return the index in which we store RDF {@link Value}s.
*
* @param jnl
* The index manager.
*
* @return The index.
*/
/*
* TODO CanonicalHuffmanRabaCoder for U1 drops the average leaf size
*
* @ m=512 from 24k to 16k. Experiment with performance tradeoff
* when compared with gzip of the record.
*
* No apparent impact for U1 on the leaves or nodes for 32 versus 8
* on the front-coded raba.
*
* Dropping maxRecLen from 256 to 64 reduces the leaves from 16k to
* 10k. Dropping it to ZERO (0) reduces the leaves to 5k. This
* suggests that we could to much better if we keep all RDF Values
* out of the index. In standalone, we can give people a TermId
* which is the raw record address. However, in scale-out it needs
* to be the key (to locate the shard) and we will resolve the RDF
* Value using the index on the shard.
*
* Suffix compression would allow us to generalize the counter and
* avoid index space costs when collisions are rare while being able
* to tolerate more collisions (short versus byte).
U1: m=800, q=8000, ratio=8, maxRecLen=0,
Elapsed: 41340ms
NumStatements: 1000313
NumDistinctVals: 291259
TotalKeyBytes: 1747554
TotalValBytes: 60824514
MaxCollisions: 1
TotalCollisions: 6
Journal size: 209715200 bytes
Average node: 9813
Average leaf: 6543
U1: m=800, q=8000, ratio=32, maxRecLen=0,
Elapsed: 40971ms
NumStatements: 1000313
NumDistinctVals: 291259
TotalKeyBytes: 1747554
TotalValBytes: 60824514
MaxCollisions: 1
TotalCollisions: 6
Journal size: 209715200 bytes
Average node: 9821
Average leaf: 6478
U1: m=800, q=8000, ratio=64, maxRecLen=0,
Elapsed: 41629ms
NumStatements: 1000313
NumDistinctVals: 291259
TotalKeyBytes: 1747554
TotalValBytes: 60824514
MaxCollisions: 1
TotalCollisions: 6
Journal size: 209715200 bytes
Average node: 9822
Average leaf: 6467
U1: m=512, q=8000, ratio=32, maxRecLen=0,
Elapsed: 44722ms
NumStatements: 1000313
NumDistinctVals: 291259
TotalKeyBytes: 1747554
TotalValBytes: 60824514
MaxCollisions: 1
TotalCollisions: 6
Journal size: 209715200 bytes
Average node/leaf: 3969 4149
U1: m=512, q=8000, ratio=32, maxRecLen=0,
Elapsed: 40519ms
NumStatements: 1000313
NumDistinctVals: 291259
TotalKeyBytes: 1747554
TotalValBytes: 60824514
MaxCollisions: 1
TotalCollisions: 6
Journal size: 209715200 bytes
Average node/leaf, node(min/max), leaf(min/max): 7583 8326 7583 7583 5755 14660
It would be great if we tracked the node/leaf data live on the RWStore for
these counters so it could all be reported periodically (via http) or at the
end in a summary.
TODO The front compression of the keys is not helping out much since the keys
are so sparse in the hash code space. It is a Good Thing that the keys are so
sparse, but this suggests that we should try a different coder for the leaf keys.
*/
private BTree getTermsIndex(final Journal jnl) {
final String name = "TERMS";
BTree ndx = jnl.getIndex(name);
final int m = 1024;
final int q = 8000;
final int ratio = 32;
final int maxRecLen = 0;
if(ndx == null) {
final IndexMetadata md = new IndexMetadata(name, UUID.randomUUID());
md.setNodeKeySerializer(new FrontCodedRabaCoder(ratio));
final DefaultTupleSerializer tupleSer = new DefaultTupleSerializer(
new DefaultKeyBuilderFactory(new Properties()),//
/*
* leaf keys
*/
// DefaultFrontCodedRabaCoder.INSTANCE,//
new FrontCodedRabaCoder(ratio),//
// CanonicalHuffmanRabaCoder.INSTANCE,
/*
* leaf values
*/
CanonicalHuffmanRabaCoder.INSTANCE
// new SimpleRabaCoder()//
);
md.setTupleSerializer(tupleSer);
// enable raw record support.
md.setRawRecords(true);
// set the maximum length of a byte[] value in a leaf.
md.setMaxRecLen(maxRecLen);
/*
* increase the branching factor since leaf size is smaller w/o
* large records.
*/
md.setBranchingFactor(m);
// Note: You need to give sufficient heap for this option!
md.setWriteRetentionQueueCapacity(q);
ndx = jnl.registerIndex(name, md);
}
return ndx;
}
private void parseFileOrDirectory(final File fileOrDir,
final RDFFormat fallback) throws Exception {
if (fileOrDir.isDirectory()) {
final File[] files = fileOrDir.listFiles();
for (int i = 0; i < files.length; i++) {
final File f = files[i];
parseFileOrDirectory(f, fallback);
}
return;
}
final File f = fileOrDir;
final String n = f.getName();
RDFFormat fmt = RDFFormat.forFileName(n, fallback);
if (fmt == null && n.endsWith(".zip")) {
fmt = RDFFormat.forFileName(n.substring(0, n.length() - 4),
fallback);
}
if (fmt == null && n.endsWith(".gz")) {
fmt = RDFFormat.forFileName(n.substring(0, n.length() - 3),
fallback);
}
if (fmt == null) {
log.warn("Ignoring: " + f);
return;
}
final StatementHandler stmtHandler = new StatementHandler(valBufSize,
c, conf, vf, mmgr, valueQueue, parsing);
final FutureTask ft = new ReportingFutureTask(
f,
new ParseFileTask(f, fallback, fileBufSize, vf, stmtHandler)
);
if (parserService != null) {
// run on the thread pool.
parserService.submit(ft);
} else {
// Run in the caller's thread.
ft.run();
// Test the Future.
ft.get();
}
}
/**
* Task parses a single file.
*
* @author thompsonbry
*/
private static class ParseFileTask implements Callable {
private final File file;
private final RDFFormat fallback;
private final int fileBufSize;
private final BigdataValueFactory vf;
private final StatementHandler stmtHandler;
public ParseFileTask(final File file, final RDFFormat fallback,
final int fileBufSize, final BigdataValueFactory vf,
final StatementHandler stmtHandler) {
if (file == null)
throw new IllegalArgumentException();
if (stmtHandler == null)
throw new IllegalArgumentException();
this.file = file;
this.fallback = fallback;
this.fileBufSize = fileBufSize;
this.vf = vf;
this.stmtHandler = stmtHandler;
}
public Void call() throws Exception {
parseFile(file);
return (Void) null;
}
private void parseFile(final File file) throws IOException,
RDFParseException, RDFHandlerException,
NoSuchAlgorithmException, InterruptedException {
if (!file.exists())
throw new RuntimeException("Not found: " + file);
final RDFFormat format = RDFFormat.forFileName(file.getName(),fallback);
if (format == null)
throw new RuntimeException("Unknown format: " + file);
if (log.isTraceEnabled())
log.trace("RDFFormat=" + format);
final RDFParserFactory rdfParserFactory = RDFParserRegistry
.getInstance().get(format);
if (rdfParserFactory == null)
throw new RuntimeException("No parser for format: " + format);
final RDFParser rdfParser = rdfParserFactory.getParser();
rdfParser.setValueFactory(vf);
rdfParser.setVerifyData(false);
rdfParser.setStopAtFirstError(false);
rdfParser.setDatatypeHandling(RDFParser.DatatypeHandling.IGNORE);
rdfParser.setRDFHandler(stmtHandler);
/*
* Run the parser, which will cause statements to be inserted.
*/
if (log.isDebugEnabled())
log.debug("Parsing: " + file);
InputStream is = new FileInputStream(file);
try {
is = new BufferedInputStream(is, fileBufSize);
final boolean gzip = file.getName().endsWith(".gz");
if (gzip)
is = new GZIPInputStream(is);
final String baseURI = file.toURI().toString();
// parse the file
rdfParser.parse(is, baseURI);
} finally {
is.close();
}
}
}
/**
* Helper class adds statements to the sail as they are visited by a parser.
*/
static private class StatementHandler extends RDFHandlerBase {
// private static final transient Logger log = HashCollisionUtility.log;
/**
* Various counters that we track.
*/
private final Counters c;
/** The lexicon configuration. */
private final LexiconConfiguration conf;
/**
* Blocking queue to which we add {@link ValueBuffer} instances as they
* are generated by the parser.
*/
final BlockingQueue valueQueue;
/**
* true iff the parser is permitted to run and
* false if the parser should terminate.
*/
final AtomicBoolean parsing;
/**
* Used to build the keys (just a hash code).
*/
private final IKeyBuilder keyBuilder = KeyBuilder.newInstance();
/** Used to serialize RDF Values as byte[]s. */
private final DataOutputBuffer out = new DataOutputBuffer();
/** Used to serialize RDF Values as byte[]s. */
private final ByteArrayBuffer tbuf = new ByteArrayBuffer();
/** Used to serialize RDF Values as byte[]s. */
private final BigdataValueSerializer valSer;
/**
* Used to (de-)compress the raw values.
*
* Note: This is not thread-safe, even for decompression. You need a
* pool or thread-local instance to support concurrent reads against the
* TERMS index.
*/
private final RecordCompressor compressor = new RecordCompressor(
Deflater.BEST_SPEED);
// /** Buffer for (serialized) RDF Values. */
// private KV[] values;
/** #of buffered values. */
private int nvalues = 0;
/** The memory manager. */
private final IMemoryManager memoryManager;
/** The current allocation context. */
private IMemoryManager context = null;
/**
* Map of distinct values in the buffer.
*
* TODO In addition to enforcing DISTINCT over the Values in the
* ValueBuffer, an LRU/LIRS cache would be nice here so we can reuse the
* frequently resolved (BigdataValue => IV) mappings across buffer
* instances.
*
* FIXME We need to provide a canonicalizing mapping for blank nodes.
*
* TODO The key should also include the URI,Literal,BNode, etc. prefix
* bits (or is this necessary any more?).
*/
private Map addrMap;
/** The size of the {@link #values} buffer when it is allocated. */
private final int valueBufSize;
public StatementHandler(//
final int valueBufSize,
final Counters c,
final LexiconConfiguration conf,
final BigdataValueFactory vf,
final IMemoryManager memoryManager,
final BlockingQueue valueQueue,
final AtomicBoolean parsing) {
this.valueBufSize = valueBufSize;
this.c = c;
this.conf = conf;
this.memoryManager = memoryManager;
this.valueQueue = valueQueue;
this.parsing = parsing;
this.valSer = vf.getValueSerializer();
}
public void endRDF() {
if(log.isTraceEnabled())
log.trace("End of source.");
try {
flush();
} catch (InterruptedException e) {
throw new RuntimeException(e);
}
}
public void handleStatement(final Statement stmt)
throws RDFHandlerException {
if (!parsing.get()) {
// Either shutdown or never started.
throw new IllegalStateException();
}
try {
bufferValue((BigdataValue) stmt.getSubject());
bufferValue((BigdataValue) stmt.getPredicate());
bufferValue((BigdataValue) stmt.getObject());
if (stmt.getContext() != null) {
bufferValue((BigdataValue) stmt.getContext());
}
} catch (InterruptedException ex) {
// Interrupted while blocked on the valueQueue
throw new RDFHandlerException(ex);
}
c.nstmts.incrementAndGet();
}
/**
* If the RDF {@link Value} can not be represented inline within the
* statement indices, then buffer the value for batch resolution against
* the TERMS index.
*
* @param value
* The RDF {@link Value}.
*
* @return A {@link Value}. If the caller's {@link Value} could be
* represented as an inline {@link IV}, then the returned value
* will be a {@link BigdataValue} and the inline {@link IV} will
* be available from {@link BigdataValue#getIV()}. Otherwise the
* caller's {@link Value} is returned and the {@link Value} must
* be resolved against the TERMS index in order to obtain its
* {@link IV}.
*
* @throws InterruptedException
*
* FIXME Handle {@link BlobIV}, {@link PartlyInlineURIIV}, and
* {@link PartlyInlineTypedLiteralIV}. These are three kinds of
* "non-inline" values. They will have to be queued for
* insertion into the TERMS index and Statement which depend
* on those non-inline values will have to be deferred until
* we have resolved those non-inline values. This is
* basically the same logic that we already have for
* StatementBuffer, except that an asynchronous queue is
* being used (by this class) to do the resolution of the IV
* for large values.
*
* Other kinds of {@link IV}s which could be handled here
* would be references to large values stored in the file
* system, in S3, etc.
*/
private void bufferValue(final BigdataValue value)
throws InterruptedException {
// Not expecting the IV to already be cached.
assert value.getIV() == null;
// Attempt to inline this value.
final IV iv = conf.createInlineIV(value);
if (iv != null) {
// This is being inlined.
switch (iv.getVTE()) {
case URI:
c.nshortURIs.incrementAndGet();
break;
case BNODE:
c.nshortBNodes.incrementAndGet();
break;
case LITERAL:
c.nshortLiterals.incrementAndGet();
break;
default:
throw new AssertionError();
}
// Verify IV is cached on that Value.
assert value.getIV() == iv;
return;
}
if (context != null && context.getSlotBytes() >= valueBufSize) {
// Incremental flush of large values to the TERMS index.
flush();
}
if (context == null) {
// Lazy allocation of the buffer.
context = memoryManager.createAllocationContext();
addrMap = new LinkedHashMap();
}
/*
* Generate a key (hash code) and value (serialized and compressed)
* from the BigdataValue.
*/
final KV t = makeKV(value);
/*
* Lookup the list of addresses for RDF Values which hash to the
* same key.
*/
Bucket bucket = addrMap.get(t.key);
if (bucket == null) {
/*
* No match on that hash code key.
*/
// lay the record down on the memory manager.
final long addr = context.allocate(ByteBuffer.wrap(t.val));
// add new bucket to the map.
addrMap.put(t.key, bucket = new Bucket(t.key, addr));
nvalues++;
} else {
/*
* Either a hash collision or the value is already stored at
* a known address.
*/
{
for (Long addr : bucket.addrs) {
if (context.allocationSize(addr) != t.val.length) {
// Non-match based on the allocated record size.
continue;
}
/*
* TODO It would be more efficient to compare the data
* using the zero-copy get(addr) method.
*/
final byte[] tmp = context.read(addr);
if (BytesUtil.bytesEqual(t.val, tmp)) {
// We've already seen this Value.
if (log.isDebugEnabled())
log.debug("Duplicate value in chunk: "
+ Arrays.toString(t.val));
/*
* FIXME This pattern does not really work out for
* building statements since we lack a reference to
* the Value which is being inserted into the TERMS
* index. The StatementBuffer handles this. It keeps
* the Values in a map and inserts all values into
* the database. [It should only keep the distinct
* non-inline values but it currently keeps all
* distinct values without regard to inlining.]
*/
return;
}
}
// Fall through - there is no such record on the store.
}
// lay the record down on the memory manager.
bucket.add(context.allocate(ByteBuffer.wrap(t.val)));
nvalues++;
}
return;
} // bufferValue()
/**
* Transfer a non-empty buffer to the {@link #valueQueue}.
*
* @throws InterruptedException
*/
void flush() throws InterruptedException {
if (nvalues == 0)
return;
if (!parsing.get()) {
// Either shutdown or never started.
throw new IllegalStateException();
}
if (log.isInfoEnabled())
log.info("Adding chunk with " + nvalues + " values and "
+ context.getUserBytes() + " bytes to queue.");
/*
* Create an object which encapsulates the allocation context (to be
* cleared when the data have been consumed) and the address map.
*/
final List contexts = new LinkedList();
contexts.add(context);
// put the buffer on the queue (blocking operation).
valueQueue.put(new ValueBuffer(contexts, nvalues, addrMap));
// clear reference since we just handed off the data.
context = null;
addrMap = null;
nvalues = 0;
// clear distinct value set so it does not build for ever.
// distinctValues.clear();
// addrMap.clear();
}
private KV makeKV(final BigdataValue r) {
byte[] val = valSer.serialize(r, out.reset(), tbuf);
/*
* FIXME In order support conditional compression we will have to
* mark the record with a header to indicate whether or not
* it is compressed. Without that header we can not
* deserialize a record resolved via its TermId since we
* will not know whether or not it is compressed (actually,
* that could be part of the termId....)
*/
if (compressor != null) {//&& val.length > 64) {
// compress, reusing [out].
out.reset();
compressor.compress(val, out);
}
// if (out.pos() < val.length) // TODO Use compressed version iff smaller.
{
val = out.toByteArray();
}
/*
* Note: This is an exclusive lower bound (it does not include the
* counter).
*
* TODO We could format the counter in here as a ZERO (0) since it
* is a fixed length value and then patch it up later. That would
* involve less copying.
*/
final byte[] key = buildKey(r, val).getKey();
return new KV(key, val);
} // makeKV()
private IKeyBuilder buildKey(final Value r, final byte[] val) {
// if (true) {
/*
* Simple 32-bit hash code based on the byte[] representation of
* the RDF Value.
*/
final int hashCode = r.hashCode();
return keyBuilder.reset().append(hashCode);
// } else {
//
// /*
// * Message digest of the serialized representation of the RDF
// * Value.
// *
// * TODO There are methods to copy out the digest (hash code)
// * without memory allocations. getDigestLength() and
// * getDigest(out,start,len).
// */
// private final MessageDigest d;
//
// try {
//
// d = MessageDigest.getInstance("SHA-256"); // 256 bits (32 bytes)
//
// } catch (NoSuchAlgorithmException e) {
//
// throw new RuntimeException(e);
//
// }
//
//
// final byte[] hashCode = d.digest(val);
//
// return keyBuilder.reset().append(hashCode);
//
// }
} // buildKey
} // class StatementHandler
/**
* Index a {@link ValueBuffer}.
*/
private static class IndexValueBufferTask implements Callable {
/**
* The {@link MemoryManager} against which the allocations were made.
*/
private final MemoryManager mmgr;
/**
* The data to be indexed.
*/
private final ValueBuffer vbuf;
/**
* The index to write on.
*/
private final BTree termsIndex;
/** Counters for things that we track. */
private final Counters c;
/**
* Used to build the keys.
*/
private final IKeyBuilder keyBuilder = KeyBuilder.newInstance();
// /** Used to serialize RDF Values as byte[]s. */
// private final DataOutputBuffer out = new DataOutputBuffer();
/** Used to de-serialize RDF Values (debugging only). */
private final BigdataValueSerializer valSer;
/**
* Used to de-compress the raw values (debugging only).
*
* Note: This is not thread-safe, even for decompression. You need a
* pool or thread-local instance to support concurrent reads against the
* TERMS index.
*/
private final RecordCompressor compressor;
public IndexValueBufferTask(final MemoryManager mmgr,
final ValueBuffer vbuf, final BTree termsIndex,
final BigdataValueFactory vf, final Counters c) {
if(mmgr == null)
throw new IllegalArgumentException();
if(vbuf == null)
throw new IllegalArgumentException();
if(termsIndex== null)
throw new IllegalArgumentException();
if(vf == null)
throw new IllegalArgumentException();
if(c == null)
throw new IllegalArgumentException();
this.mmgr = mmgr;
this.vbuf = vbuf;
this.termsIndex = termsIndex;
this.c = c;
/*
* Note: debugging only.
*/
this.valSer = vf.getValueSerializer();
this.compressor = new RecordCompressor(Deflater.BEST_SPEED);
}
public Void call() throws Exception {
final long begin = System.currentTimeMillis();
if (log.isInfoEnabled())
log.info("Indexing " + vbuf.nvalues + " values occupying "
+ vbuf.getUserBytes() + " bytes");
/*
* Place into sorted order by the keys.
*
* The Bucket implements Comparable. We extract the buckets, sort
* them, and then process them.
*/
final Bucket[] a = vbuf.addrMap.values().toArray(new Bucket[0]);
Arrays.sort(a);
// Index the values.
for (int i = 0; i = Byte.MAX_VALUE) {
/*
* Impose a hard limit on the #of hash collisions we will accept
* in this utility.
*
* @todo We do not need to have a hard limit if we use
* BigInteger for the counter, but the performance will go
* through the floor if we have to scan 32k entries on a hash
* collision!
*/
throw new RuntimeException("Too many hash collisions: ncoll="
+ rangeCount);
}
// force range count into (signed) byte
final byte counter = (byte) rangeCount;
if (rangeCount == 0) {
/*
* This is the first time we have observed a Value which
* generates this hash code, so append a [short] ZERO (0) to
* generate the actual key and then insert the Value into the
* index. Since there is nothing in the index for this hash
* code, no collision is possible and we do not need to test the
* index for the value before inserting the value into the
* index.
*/
final byte[] key = keyBuilder.reset().append(fromKey).appendSigned(
counter).getKey();
if (termsIndex.insert(key, val) != null) {
throw new AssertionError();
}
c.ninserted.incrementAndGet();
c.totalKeyBytes.addAndGet(key.length);
c.totalValBytes.addAndGet(val.length);
return;
}
/*
* iterator over that key range
*
* TODO Filter for the value of interest so we can optimize the scan
* by comparing with the value without causing it to be
* materialized, especially we should be able to efficiently reject
* tuples where the byte[] value length is known to differ from the
* a given length, including when the value is stored as a raw
* record at which point we are doing a fast rejection based on
* comparing the byteCount(addr) for the raw record with the target
* byte count for value that we are seeking in the index.
*
* We can visit something iff the desired tuple already exists (same
* length, and possibly the same data). If we visit nothing then we
* know that we have to insert a tuple and we know the counter value
* from the collision count.
*/
final ITupleIterator itr = termsIndex.rangeIterator(fromKey, toKey,
0/* capacity */, IRangeQuery.VALS, null/* filter */);
boolean found = false;
while(itr.hasNext()) {
final ITuple tuple = itr.next();
// raw bytes
final byte[] tmp = tuple.getValue();
if (false)
System.out.println(getValue(tmp));
// Note: Compares the compressed values ;-)
if(BytesUtil.bytesEqual(val, tmp)) {
found = true;
break;
}
}
if(found) {
// Already in the index.
return;
}
/*
* Hash collision.
*/
if (rangeCount > c.maxCollisions.get()) {
// Raise the maximum collision count.
c.maxCollisions.set(rangeCount);
log.warn("MAX COLLISIONS NOW: " + c.maxCollisions.get());
}
final byte[] key = keyBuilder.reset().append(fromKey).appendSigned(
counter).getKey();
// Insert into the index.
if (termsIndex.insert(key, val) != null) {
throw new AssertionError();
}
c.ninserted.incrementAndGet();
c.totalKeyBytes.addAndGet(key.length);
c.totalValBytes.addAndGet(val.length);
c.totalCollisions.incrementAndGet();
if (rangeCount > 128) { // arbitrary limit to log @ WARN.
log.warn("Collision: hashCode=" + BytesUtil.toString(key)
+ ", nstmts="+c.nstmts
+ ", nshortLiterals=" + c.nshortLiterals
+ ", nshortURIs=" + c.nshortURIs + ", ninserted="
+ c.ninserted + ", totalCollisions=" + c.totalCollisions
+ ", maxCollisions=" + c.maxCollisions
+ ", ncollThisTerm=" + rangeCount + ", resource="
+ getValue(val));
} else if (log.isDebugEnabled())
log.debug("Collision: hashCode=" + BytesUtil.toString(key)
+ ", nstmts="+c.nstmts
+ ", nshortLiterals=" + c.nshortLiterals
+ ", nshortURIs=" + c.nshortURIs + ", ninserted="
+ c.ninserted + ", totalCollisions=" + c.totalCollisions
+ ", maxCollisions=" + c.maxCollisions
+ ", ncollThisTerm=" + rangeCount + ", resource="
+ getValue(val));
}
/**
* Decompress and deserialize a {@link Value}.
*
* @param tmp
* The serialized and compressed value.
*
* @return The {@link Value}.
*/
private Value getValue(final byte[] tmp) {
// decompress
final ByteBuffer b = compressor.decompress(tmp);
final byte[] c = new byte[b.limit()];
b.get(c);
// deserialize.
return valSer.deserialize(c);
}
} // class IndexValueBufferTask
/**
* Parse files, inserting {@link Value}s into indices and counting hash
* collisions.
*
* @param args
* filename(s)
*
* @throws IOException
* @throws RDFHandlerException
* @throws RDFParseException
* @throws NoSuchAlgorithmException
*/
public static void main(final String[] args) throws Exception {
Banner.banner();
// check args.
{
for (String filename : args) {
final File file = new File(filename);
if (!file.exists())
throw new RuntimeException("Not found: " + file);
}
}
final long begin = System.currentTimeMillis();
final Properties properties = new Properties();
properties.setProperty(Journal.Options.BUFFER_MODE, BufferMode.DiskRW
.toString());
properties.setProperty(Journal.Options.INITIAL_EXTENT, ""
+ (Bytes.megabyte * 200));
// properties.setProperty(Journal.Options.COLLECT_PLATFORM_STATISTICS,"true");
// properties.setProperty(Journal.Options.COLLECT_QUEUE_STATISTICS,"true");
properties.setProperty(Journal.Options.HTTPD_PORT,"8081");
// The caller MUST specify the filename using -D on the command line.
final String journalFile = System.getProperty(Journal.Options.FILE);
if (journalFile == null) {
System.err.println("Journal file must be specified: -D"
+ Journal.Options.FILE);
System.exit(1);
}
properties.setProperty(Journal.Options.FILE, journalFile);
if (new File(journalFile).exists()) {
System.err.println("Removing old journal: " + journalFile);
new File(journalFile).delete();
}
final Journal jnl = new Journal(properties);
final RDFFormat fallback = RDFFormat.N3;
HashCollisionUtility u = null;
try {
u = new HashCollisionUtility(jnl);
u.start();
for (String filename : args) {
u.parseFileOrDirectory(new File(filename), fallback);
}
// // flush anything left in the buffer.
// u.stmtHandler.flush();
// shutdown and block until all data is indexed.
u.shutdown();
jnl.commit();
} catch (Throwable t) {
u.shutdownNow();
throw new RuntimeException(t);
} finally {
jnl.close();
final long elapsed = System.currentTimeMillis() - begin;
System.out.println("Elapsed: " + elapsed + "ms");
if (u != null) {
System.out.println("NumStatements: " + u.c.nstmts);
System.out.println("NumDistinctVals: " + u.c.ninserted);
System.out.println("NumShortLiterals: " + u.c.nshortLiterals);
System.out.println("NumShortBNodes: " + u.c.nshortBNodes);
System.out.println("NumShortURIs: " + u.c.nshortURIs);
// System.out.println("NumCacheHit: " + u.ncached);
System.out.println("TotalKeyBytes: " + u.c.totalKeyBytes);
System.out.println("TotalValBytes: " + u.c.totalValBytes);
System.out.println("MaxCollisions: " + u.c.maxCollisions);
System.out.println("TotalCollisions: " + u.c.totalCollisions);
}
if (new File(journalFile).exists()) {
System.out.println("Journal size: "
+ new File(journalFile).length() + " bytes");
}
}
}
}