com.bigdata.journal.ConcurrencyManager Maven / Gradle / Ivy
/**
Copyright (C) SYSTAP, LLC DBA Blazegraph 2006-2016. All rights reserved.
Contact:
SYSTAP, LLC DBA Blazegraph
2501 Calvert ST NW #106
Washington, DC 20008
[email protected]
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; version 2 of the License.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*/
/*
* Created on Oct 10, 2007
*/
package com.bigdata.journal;
import java.util.Collection;
import java.util.LinkedList;
import java.util.List;
import java.util.Properties;
import java.util.concurrent.ArrayBlockingQueue;
import java.util.concurrent.BlockingQueue;
import java.util.concurrent.CancellationException;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.Future;
import java.util.concurrent.FutureTask;
import java.util.concurrent.LinkedBlockingQueue;
import java.util.concurrent.RejectedExecutionException;
import java.util.concurrent.ScheduledExecutorService;
import java.util.concurrent.SynchronousQueue;
import java.util.concurrent.ThreadPoolExecutor;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.TimeoutException;
import org.apache.log4j.Logger;
import com.bigdata.BigdataStatics;
import com.bigdata.btree.BTree;
import com.bigdata.concurrent.NonBlockingLockManager;
import com.bigdata.concurrent.NonBlockingLockManagerWithNewDesign;
import com.bigdata.counters.CounterSet;
import com.bigdata.counters.ICounterSet;
import com.bigdata.counters.Instrument;
import com.bigdata.resources.StoreManager;
import com.bigdata.service.AbstractDistributedFederation;
import com.bigdata.service.IBigdataClient;
import com.bigdata.service.IServiceShutdown;
import com.bigdata.util.DaemonThreadFactory;
import com.bigdata.util.concurrent.TaskCounters;
import com.bigdata.util.concurrent.ThreadPoolExecutorStatisticsTask;
import com.bigdata.util.concurrent.WriteTaskCounters;
/**
* Supports concurrent operations against named indices. Historical read and
* read-committed tasks run with full concurrency. For unisolated tasks, the
* {@link ConcurrencyManager} uses a {@link NonBlockingLockManager} to identify
* a schedule of operations such that access to an unisolated named index is
* always single threaded while access to distinct unisolated named indices MAY
* be concurrent.
*
* There are several thread pools that facilitate concurrency. They are:
*
*
* - {@link #readService}
* - Concurrent historical and read-committed tasks are run against a
* historical view of a named index using this service. No
* locking is imposed. Concurrency is limited by the size of the thread pool.
*
* - {@link #writeService}
* - Concurrent unisolated writers running against the current
* view of (or more more) named index(s) (the "live" or "mutable" index(s)). The
* underlying {@link BTree} is NOT thread-safe for writers. Therefore writers
* MUST predeclare their locks, which allows us to avoid deadlocks altogether.
* This is also used to schedule the commit phrase of transactions (transaction
* commits are in fact unisolated tasks).
*
* - {@link #txWriteService}
* -
*
* This is used for the "active" phrase of transaction. Transactions read from
* historical states of named indices during their active phase and buffer the
* results on isolated indices backed by a per-transaction
* {@link TemporaryStore}. Since transactions never write on the unisolated
* indices during their "active" phase, distinct transactions may be run with
* arbitrary concurrency. However, concurrent tasks for the same transaction
* must obtain an exclusive lock on the isolated index(s) that are used to
* buffer their writes.
*
*
* A transaction that requests a commit using the
* {@link ITransactionManagerService} results in a unisolated task being
* submitted to the {@link #writeService}. Transactions are selected to commit
* once they have acquired a lock on the corresponding unisolated indices,
* thereby enforcing serialization of their write sets both among other
* transactions and among unisolated writers. The commit itself consists of the
* standard validation and merge phrases.
*
*
*
* @author Bryan Thompson
*/
public class ConcurrencyManager implements IConcurrencyManager {
static final private Logger log = Logger.getLogger(ConcurrencyManager.class);
// /**
// * True iff the {@link #log} level is INFO or less.
// */
// final protected static boolean INFO = log.isInfoEnabled();
//
// /**
// * True iff the {@link #log} level is DEBUG or less.
// */
// final private static boolean DEBUG = log.isDebugEnabled();
/**
* Options for the {@link ConcurrentManager}.
*
* @author Bryan Thompson
*/
public static interface Options extends IServiceShutdown.Options {
/**
* The #of threads in the pool handling concurrent transactions.
*
* @see #DEFAULT_TX_SERVICE_CORE_POOL_SIZE
*/
String TX_SERVICE_CORE_POOL_SIZE = ConcurrencyManager.class.getName()
+ ".txService.corePoolSize";
/**
* The default #of threads in the transaction service thread pool.
*/
String DEFAULT_TX_SERVICE_CORE_POOL_SIZE = "0";
/**
* The #of threads in the pool handling concurrent unisolated read
* requests on named indices -or- ZERO (0) if the size of the thread
* pool is not fixed (default is 0).
*
* @see #DEFAULT_READ_SERVICE_CORE_POOL_SIZE
*/
String READ_SERVICE_CORE_POOL_SIZE = ConcurrencyManager.class.getName()
+ ".readService.corePoolSize";
/**
* The default #of threads in the read service thread pool.
*
* @see #READ_SERVICE_CORE_POOL_SIZE
*/
String DEFAULT_READ_SERVICE_CORE_POOL_SIZE = "0";
/**
* The minimum #of threads in the pool handling concurrent unisolated
* write on named indices (default is
* {@value #DEFAULT_WRITE_SERVICE_CORE_POOL_SIZE}). The size of the
* thread pool will automatically grow to meet the possible concurrency
* of the submitted tasks up to the configured
* {@link #WRITE_SERVICE_MAXIMUM_POOL_SIZE}.
*
* The main factor that influences the throughput of group commit is the
* potential concurrency of the submitted
* {@link ITx#UNISOLATED} tasks (both how many can be submitted in
* parallel by the application and how many can run concurrency - tasks
* writing on the same index have a shared dependency and can not run
* concurrently).
*
* Each {@link ITx#UNISOLATED} task that completes processing will block
* until the next commit, thereby absorbing a worker thread. For this
* reason, it can improve performance if you set the core pool size and
* the maximum pool size for the write service higher that
* the potential concurrency with which the submitted tasks could be
* processed.
*
* There is also a strong effect as the JVM performs optimizations on
* the running code, so randomize your tests. See
* {@link StressTestGroupCommit} for performance tuning.
*
* @see #DEFAULT_WRITE_SERVICE_CORE_POOL_SIZE
*/
String WRITE_SERVICE_CORE_POOL_SIZE = ConcurrencyManager.class
.getName()
+ ".writeService.corePoolSize";
/**
* The default minimum #of threads in the write service thread pool.
*/
String DEFAULT_WRITE_SERVICE_CORE_POOL_SIZE = "10";
/**
* The maximum #of threads allowed in the pool handling concurrent
* unisolated write on named indices (default is
* {@value #DEFAULT_WRITE_SERVICE_CORE_POOL_SIZE}.
*
* Note: This property is ignored if the
* {@link #WRITE_SERVICE_QUEUE_CAPACITY} is ZERO (0) or
* {@link Integer#MAX_VALUE}!
*
* @see #DEFAULT_WRITE_SERVICE_MAXIMUM_POOL_SIZE
*/
String WRITE_SERVICE_MAXIMUM_POOL_SIZE = ConcurrencyManager.class
.getName()
+ ".writeService.maximumPoolSize";
/**
* The default for the maximum #of threads in the write service thread
* pool.
*/
String DEFAULT_WRITE_SERVICE_MAXIMUM_POOL_SIZE = "50";
/**
* The time in milliseconds that the {@link WriteExecutorService} will
* keep alive excess worker threads (those beyond the core pool size).
*/
String WRITE_SERVICE_KEEP_ALIVE_TIME = ConcurrencyManager.class
.getName()
+ ".writeService.keepAliveTime";
String DEFAULT_WRITE_SERVICE_KEEP_ALIVE_TIME = "60000";
/**
* When true, the write service will be prestart all of its worker
* threads (default
* {@value #DEFAULT_WRITE_SERVICE_PRESTART_ALL_CORE_THREADS}).
*
* @see #DEFAULT_WRITE_SERVICE_PRESTART_ALL_CORE_THREADS
*/
String WRITE_SERVICE_PRESTART_ALL_CORE_THREADS = ConcurrencyManager.class
.getName()
+ ".writeService.prestartAllCoreThreads";
/**
* The default for {@link #WRITE_SERVICE_PRESTART_ALL_CORE_THREADS}.
*/
String DEFAULT_WRITE_SERVICE_PRESTART_ALL_CORE_THREADS = "false";
/**
* The maximum capacity of the write service queue before newly
* submitted tasks will be rejected -or- ZERO (0) to use a
* {@link SynchronousQueue} (default
* {@value.#DEFAULT_WRITE_SERVICE_SYNCHRONOUS_QUEUE_CAPACITY}).
*
* Note: When the {@link #WRITE_SERVICE_QUEUE_CAPACITY} is ZERO (0), a
* {@link SynchronousQueue} is used, the maximumPoolSize is ignored, and
* new {@link Thread}s will be created on demand. This allow the #of
* threads to change in response to demand while ensuring that tasks are
* never rejected.
*
* Note: A {@link LinkedBlockingQueue} will be used if the queue
* capacity is {@link Integer#MAX_VALUE}. The corePoolSize will never
* increase for an unbounded queue so the value specified for
* maximumPoolSize will be ignored and tasks will never be rejected. See
* {@link ThreadPoolExecutor}'s discussion on queues for more
* information on this issue.
*
* Note: When a bounded queue capacity is specified, tasks will be
* rejected if the the corePoolThreads are busy and the work queue is
* full.
*
* @see ThreadPoolExecutor
* @see #DEFAULT_WRITE_SERVICE_QUEUE_CAPACITY
*/
String WRITE_SERVICE_QUEUE_CAPACITY = ConcurrencyManager.class
.getName()
+ ".writeService.queueCapacity";
/**
* The default maximum depth of the write service queue (0).
*/
String DEFAULT_WRITE_SERVICE_QUEUE_CAPACITY = "0";
/**
* The timeout in milliseconds that the the {@link WriteExecutorService}
* will await other tasks to join the commit group (default
* {@value #DEFAULT_WRITE_SERVICE_GROUP_COMMIT_TIMEOUT}). When ZERO
* (0), group commit is disabled since the first task to join the commit
* group will NOT wait for other tasks and the commit group will
* therefore always consist of a single task.
*
* @see #DEFAULT_WRITE_SERVICE_GROUP_COMMIT_TIMEOUT
*/
String WRITE_SERVICE_GROUP_COMMIT_TIMEOUT = ConcurrencyManager.class
.getName()
+ ".writeService.groupCommitTimeout";
String DEFAULT_WRITE_SERVICE_GROUP_COMMIT_TIMEOUT = "100";
/**
* The time in milliseconds that a group commit will await an exclusive
* lock on the write service in order to perform synchronous overflow
* processing (default
* {@value #DEFAULT_WRITE_SERVICE_OVERFLOW_LOCK_REQUEST_TIMEOUT}). This
* lock is requested IFF overflow process SHOULD be performed
* (asynchronous overflow processing is enabled, asynchronous overflow
* processing is not ongoing, and the live journal extent exceeds the
* threshold extent).
*
* The lock timeout needs to be of significant duration or a lock
* request for a write service under heavy write load will timeout, in
* which case an error will be logged. If overflow processing is not
* performed the live journal extent will grow without bound and the
* service will be unable to release older resources on the disk.
*/
String WRITE_SERVICE_OVERFLOW_LOCK_REQUEST_TIMEOUT = ConcurrencyManager.class
.getName()
+ ".writeService.overflowLockRequestTimeout";
String DEFAULT_WRITE_SERVICE_OVERFLOW_LOCK_REQUEST_TIMEOUT = ""
+ (2 * 60 * 1000);
}
/**
* The properties specified to the ctor.
*/
final private Properties properties;
/**
* The object managing local transactions.
*/
final private ILocalTransactionManager transactionManager;
/**
* The object managing the resources on which the indices are stored.
*/
final private IResourceManager resourceManager;
/**
* The local time at which this service was started.
*/
final private long serviceStartTime = System.currentTimeMillis();
/**
* true until the service is shutdown.
*/
private volatile boolean open = true;
/**
* Pool of threads for handling concurrent read/write transactions on named
* indices. Distinct transactions are not inherently limited in their
* concurrency, but concurrent operations within a single transaction MUST
* obtain an exclusive lock on the isolated index(s) on the temporary store.
* The size of the thread pool for this service governs the maximum
* practical concurrency for transactions.
*
* Transactions always read from historical data and buffer their writes
* until they commit. Transactions that commit MUST acquire unisolated
* writable indices for each index on which the transaction has written.
* Once the transaction has acquired those writable indices it then runs its
* commit phrase as an unisolated operation on the {@link #writeService}.
*/
final private ThreadPoolExecutor txWriteService;
/**
* Pool of threads for handling concurrent unisolated read operations on
* named indices using historical data. Unisolated
* read operations from historical data are not inherently limited in
* their concurrency and do not conflict with unisolated writers. The
* size of the thread pool for this service governs the maximum
* practical concurrency for unisolated readers.
*
* Note that unisolated read operations on the current
* state of an index DO conflict with unisolated writes and such tasks
* must be run as unisolated writers.
*
* Note: unisolated readers of historical data do require the rention of
* historical commit records (which may span more than one logical
* journal) until the reader terminates.
*/
final private ThreadPoolExecutor readService;
/**
* Pool of threads for handling concurrent unisolated write operations on
* named indices and namespaces spanning multiple named indices (via
* hierarchical locking). Unisolated writes are always performed against the
* current state of the named index. Unisolated writes for the same named
* index (or index partition) conflict and must be serialized. The size of
* this thread pool and the #of distinct named indices (or namespaces)
* together govern the maximum practical concurrency for unisolated writers.
*
* Serialization of access to unisolated named indices is accomplished by
* gaining an exclusive lock on the named resource(s) corresponding to the
* index(es) or namespace.
*/
private final WriteExecutorService writeService;
/**
* When true the {@link #sampleService} will be used run
* {@link ThreadPoolExecutorStatisticsTask}s that collect statistics on the
* {@link #readService}, {@link #writeService}, and the
* {@link #txWriteService}.
*/
final private boolean collectQueueStatistics;
/**
* Used to sample some counters at a once-per-second rate.
*/
final private ScheduledExecutorService sampleService;
/**
* The timeout for {@link #shutdown()} -or- ZERO (0L) to wait for ever.
*/
final private long shutdownTimeout;
// /**
// * An object wrapping the properties specified to the ctor.
// */
// public Properties getProperties() {
//
// return new Properties(properties);
//
// }
private void assertOpen() {
if (!open)
throw new IllegalStateException();
}
@Override
public WriteExecutorService getWriteService() {
assertOpen();
return writeService;
}
// public LockManager getLockManager() {
//
// assertOpen();
//
// return lockManager;
//
// }
@Override
public ILocalTransactionManager getTransactionManager() {
assertOpen();
return transactionManager;
}
@Override
public IResourceManager getResourceManager() {
assertOpen();
return resourceManager;
}
@Override
public boolean isOpen() {
return open;
}
/**
* Shutdown the thread pools (running tasks will run to completion, but no
* new tasks will start).
*/
@Override
synchronized public void shutdown() {
if(!isOpen()) return;
open = false;
if (log.isInfoEnabled())
log.info("begin");
// time when shutdown begins.
final long begin = System.currentTimeMillis();
/*
* Note: when the timeout is zero we approximate "forever" using
* Long.MAX_VALUE.
*/
final long shutdownTimeout = this.shutdownTimeout == 0L ? Long.MAX_VALUE
: this.shutdownTimeout;
final TimeUnit unit = TimeUnit.MILLISECONDS;
txWriteService.shutdown();
readService.shutdown();
writeService.shutdown();
if (sampleService != null)
sampleService.shutdown();
try {
if (log.isInfoEnabled()) log.info("Awaiting transaction service termination");
final long elapsed = System.currentTimeMillis() - begin;
if(!txWriteService.awaitTermination(shutdownTimeout-elapsed, unit)) {
log.warn("Transaction service termination: timeout");
}
} catch(InterruptedException ex) {
log.warn("Interrupted awaiting transaction service termination.", ex);
}
try {
if (log.isInfoEnabled())
log.info("Awaiting read service termination");
final long elapsed = System.currentTimeMillis() - begin;
if(!readService.awaitTermination(shutdownTimeout-elapsed, unit)) {
log.warn("Read service termination: timeout");
}
} catch(InterruptedException ex) {
log.warn("Interrupted awaiting read service termination.", ex);
}
try {
final long elapsed = System.currentTimeMillis() - begin;
final long timeout = shutdownTimeout-elapsed;
if (log.isInfoEnabled())
log.info("Awaiting write service termination: will wait "
+ timeout + "ms");
if(!writeService.awaitTermination(timeout, unit)) {
log.warn("Write service termination : timeout");
}
} catch(InterruptedException ex) {
log.warn("Interrupted awaiting write service termination.", ex);
}
final long elapsed = System.currentTimeMillis() - begin;
if (log.isInfoEnabled())
log.info("Done: elapsed=" + elapsed + "ms");
}
/**
* Immediate shutdown (running tasks are canceled rather than being
* permitted to complete).
*
* @see #shutdown()
*/
@Override
public void shutdownNow() {
if(!isOpen()) return;
open = false;
if (log.isInfoEnabled())
log.info("begin");
final long begin = System.currentTimeMillis();
txWriteService.shutdownNow();
readService.shutdownNow();
writeService.shutdownNow();
if (sampleService != null)
sampleService.shutdownNow();
final long elapsed = System.currentTimeMillis() - begin;
if (log.isInfoEnabled())
log.info("Done: elapsed=" + elapsed + "ms");
}
/**
* (Re-)open a journal supporting concurrent operations.
*
* @param properties
* See {@link ConcurrencyManager.Options}.
* @param transactionManager
* The object managing the local transactions.
* @param resourceManager
* The object managing the resources on which the indices are
* stored.
*/
public ConcurrencyManager(final Properties properties,
final ILocalTransactionManager transactionManager,
final IResourceManager resourceManager) {
if (properties == null)
throw new IllegalArgumentException();
if (transactionManager == null)
throw new IllegalArgumentException();
if (resourceManager == null)
throw new IllegalArgumentException();
this.properties = properties;
this.transactionManager = transactionManager;
this.resourceManager = resourceManager;
String val;
final int txServicePoolSize;
final int readServicePoolSize;
// txServicePoolSize
{
val = properties.getProperty(ConcurrencyManager.Options.TX_SERVICE_CORE_POOL_SIZE,
ConcurrencyManager.Options.DEFAULT_TX_SERVICE_CORE_POOL_SIZE);
txServicePoolSize = Integer.parseInt(val);
if (txServicePoolSize < 0) {
throw new RuntimeException("The '"
+ ConcurrencyManager.Options.TX_SERVICE_CORE_POOL_SIZE
+ "' must be non-negative.");
}
if (log.isInfoEnabled())
log.info(ConcurrencyManager.Options.TX_SERVICE_CORE_POOL_SIZE
+ "=" + txServicePoolSize);
}
// readServicePoolSize
{
val = properties.getProperty(ConcurrencyManager.Options.READ_SERVICE_CORE_POOL_SIZE,
ConcurrencyManager.Options.DEFAULT_READ_SERVICE_CORE_POOL_SIZE);
readServicePoolSize = Integer.parseInt(val);
if (readServicePoolSize < 0) {
throw new RuntimeException("The '"
+ ConcurrencyManager.Options.READ_SERVICE_CORE_POOL_SIZE
+ "' must be non-negative.");
}
if (log.isInfoEnabled())
log.info(ConcurrencyManager.Options.READ_SERVICE_CORE_POOL_SIZE
+ "=" + readServicePoolSize);
}
// shutdownTimeout
{
val = properties.getProperty(ConcurrencyManager.Options.SHUTDOWN_TIMEOUT,
ConcurrencyManager.Options.DEFAULT_SHUTDOWN_TIMEOUT);
shutdownTimeout = Long.parseLong(val);
if (shutdownTimeout < 0) {
throw new RuntimeException("The '" + ConcurrencyManager.Options.SHUTDOWN_TIMEOUT
+ "' must be non-negative.");
}
if (log.isInfoEnabled())
log.info(ConcurrencyManager.Options.SHUTDOWN_TIMEOUT + "="
+ shutdownTimeout);
}
// setup thread pool for concurrent transactions.
if (txServicePoolSize == 0) {
// cached thread pool.
txWriteService = (ThreadPoolExecutor) Executors
.newCachedThreadPool(new DaemonThreadFactory
(getClass().getName()+".txWriteService"));
} else {
// fixed thread pool.
txWriteService = (ThreadPoolExecutor) Executors.newFixedThreadPool(
txServicePoolSize, new DaemonThreadFactory
(getClass().getName()+".txWriteService"));
}
// setup thread pool for unisolated read operations.
if (readServicePoolSize == 0) {
// cached thread pool.
readService = (ThreadPoolExecutor) Executors
.newCachedThreadPool(new DaemonThreadFactory
(getClass().getName()+".readService"));
} else {
// fixed thread pool.
readService = (ThreadPoolExecutor) Executors.newFixedThreadPool(
readServicePoolSize, new DaemonThreadFactory
(getClass().getName()+".readService"));
}
// setup thread pool for unisolated write operations.
{
final int writeServiceCorePoolSize;
final int writeServiceMaximumPoolSize;
final int writeServiceQueueCapacity;
final boolean writeServicePrestart;
// writeServiceCorePoolSize
{
writeServiceCorePoolSize = Integer.parseInt(properties.getProperty(
ConcurrencyManager.Options.WRITE_SERVICE_CORE_POOL_SIZE,
ConcurrencyManager.Options.DEFAULT_WRITE_SERVICE_CORE_POOL_SIZE));
if (writeServiceCorePoolSize < 0) {
throw new RuntimeException("The '"
+ ConcurrencyManager.Options.WRITE_SERVICE_CORE_POOL_SIZE
+ "' must be non-negative.");
}
if (log.isInfoEnabled())
log.info(ConcurrencyManager.Options.WRITE_SERVICE_CORE_POOL_SIZE
+ "=" + writeServiceCorePoolSize);
}
// writeServiceMaximumPoolSize
{
writeServiceMaximumPoolSize = Integer.parseInt(properties.getProperty(
ConcurrencyManager.Options.WRITE_SERVICE_MAXIMUM_POOL_SIZE,
ConcurrencyManager.Options.DEFAULT_WRITE_SERVICE_MAXIMUM_POOL_SIZE));
if (writeServiceMaximumPoolSize < writeServiceCorePoolSize) {
throw new RuntimeException("The '"
+ ConcurrencyManager.Options.WRITE_SERVICE_MAXIMUM_POOL_SIZE
+ "' must be greater than the core pool size.");
}
if (log.isInfoEnabled())
log.info(ConcurrencyManager.Options.WRITE_SERVICE_MAXIMUM_POOL_SIZE
+ "=" + writeServiceMaximumPoolSize);
}
// writeServiceQueueCapacity
{
writeServiceQueueCapacity = Integer.parseInt(properties.getProperty(
ConcurrencyManager.Options.WRITE_SERVICE_QUEUE_CAPACITY,
ConcurrencyManager.Options.DEFAULT_WRITE_SERVICE_QUEUE_CAPACITY));
if (writeServiceQueueCapacity < 0) {
throw new RuntimeException("The '"
+ ConcurrencyManager.Options.WRITE_SERVICE_QUEUE_CAPACITY
+ "' must be non-negative.");
}
if(log.isInfoEnabled())
log.info(ConcurrencyManager.Options.WRITE_SERVICE_QUEUE_CAPACITY+ "="
+ writeServiceQueueCapacity);
}
// writeServicePrestart
{
writeServicePrestart = Boolean.parseBoolean(properties.getProperty(
ConcurrencyManager.Options.WRITE_SERVICE_PRESTART_ALL_CORE_THREADS,
ConcurrencyManager.Options.DEFAULT_WRITE_SERVICE_PRESTART_ALL_CORE_THREADS));
if (log.isInfoEnabled())
log.info(ConcurrencyManager.Options.WRITE_SERVICE_PRESTART_ALL_CORE_THREADS
+ "=" + writeServicePrestart);
}
final long groupCommitTimeout = Long
.parseLong(properties
.getProperty(
ConcurrencyManager.Options.WRITE_SERVICE_GROUP_COMMIT_TIMEOUT,
ConcurrencyManager.Options.DEFAULT_WRITE_SERVICE_GROUP_COMMIT_TIMEOUT));
if (log.isInfoEnabled())
log
.info(ConcurrencyManager.Options.WRITE_SERVICE_GROUP_COMMIT_TIMEOUT
+ "=" + groupCommitTimeout);
final long overflowLockRequestTimeout = Long
.parseLong(properties
.getProperty(
ConcurrencyManager.Options.WRITE_SERVICE_OVERFLOW_LOCK_REQUEST_TIMEOUT,
ConcurrencyManager.Options.DEFAULT_WRITE_SERVICE_OVERFLOW_LOCK_REQUEST_TIMEOUT));
if (log.isInfoEnabled())
log
.info(ConcurrencyManager.Options.WRITE_SERVICE_OVERFLOW_LOCK_REQUEST_TIMEOUT
+ "=" + overflowLockRequestTimeout);
final long keepAliveTime = Long
.parseLong(properties
.getProperty(
ConcurrencyManager.Options.WRITE_SERVICE_KEEP_ALIVE_TIME,
ConcurrencyManager.Options.DEFAULT_WRITE_SERVICE_KEEP_ALIVE_TIME));
if (log.isInfoEnabled())
log
.info(ConcurrencyManager.Options.WRITE_SERVICE_KEEP_ALIVE_TIME
+ "=" + keepAliveTime);
final boolean synchronousQueue = writeServiceQueueCapacity == 0;
final BlockingQueue queue;
if (synchronousQueue) {
queue = new SynchronousQueue();
} else if (writeServiceQueueCapacity == Integer.MAX_VALUE) {
queue = new LinkedBlockingQueue(
writeServiceQueueCapacity);
} else {
queue = new ArrayBlockingQueue(
writeServiceQueueCapacity);
}
writeService = new WriteExecutorService(//
resourceManager,//
writeServiceCorePoolSize,//
synchronousQueue?Integer.MAX_VALUE:writeServiceMaximumPoolSize,//
keepAliveTime, TimeUnit.MILLISECONDS, // keepAliveTime
queue, //
new DaemonThreadFactory(getClass().getName()+".writeService"), //
groupCommitTimeout,//
overflowLockRequestTimeout
);
if (writeServicePrestart) {
getWriteService().prestartAllCoreThreads();
}
}
{
collectQueueStatistics = Boolean
.parseBoolean(properties
.getProperty(
IBigdataClient.Options.COLLECT_QUEUE_STATISTICS,
IBigdataClient.Options.DEFAULT_COLLECT_QUEUE_STATISTICS));
if (log.isInfoEnabled())
log.info(IBigdataClient.Options.COLLECT_QUEUE_STATISTICS + "="
+ collectQueueStatistics);
}
if (collectQueueStatistics) {
/*
* Setup once-per-second sampling for some counters.
*/
// @todo config.
final double w = ThreadPoolExecutorStatisticsTask.DEFAULT_WEIGHT;
final long initialDelay = 0; // initial delay in ms.
final long delay = 1000; // delay in ms.
final TimeUnit unit = TimeUnit.MILLISECONDS;
writeServiceQueueStatisticsTask = new ThreadPoolExecutorStatisticsTask("writeService",
getWriteService(), countersUN, w);
txWriteServiceQueueStatisticsTask = new ThreadPoolExecutorStatisticsTask("txWriteService",
txWriteService, countersTX, w);
readServiceQueueStatisticsTask = new ThreadPoolExecutorStatisticsTask("readService",
readService, countersHR, w);
sampleService = Executors.newSingleThreadScheduledExecutor(new DaemonThreadFactory
(getClass().getName()+".sampleService"));
// the write service.
sampleService.scheduleWithFixedDelay(writeServiceQueueStatisticsTask,
initialDelay, delay, unit);
// the lock service for the write service.
sampleService.scheduleWithFixedDelay(getWriteService().getLockManager().statisticsTask,
initialDelay, delay, unit);
// the tx write service.
sampleService.scheduleWithFixedDelay(txWriteServiceQueueStatisticsTask,
initialDelay, delay, unit);
// the read service.
sampleService.scheduleWithFixedDelay(readServiceQueueStatisticsTask,
initialDelay, delay, unit);
} else {
writeServiceQueueStatisticsTask = null;
txWriteServiceQueueStatisticsTask = null;
readServiceQueueStatisticsTask = null;
sampleService = null;
}
}
/** Counters for {@link #writeService}. */
protected final WriteTaskCounters countersUN = new WriteTaskCounters();
/** Counters for the {@link #txWriteService}. */
protected final TaskCounters countersTX = new TaskCounters();
/** Counters for the {@link #readService}. */
protected final TaskCounters countersHR = new TaskCounters();
/**
* Sampling instruments for the various queues giving us the moving average
* of the queue length.
*/
private final ThreadPoolExecutorStatisticsTask writeServiceQueueStatisticsTask;
private final ThreadPoolExecutorStatisticsTask txWriteServiceQueueStatisticsTask;
private final ThreadPoolExecutorStatisticsTask readServiceQueueStatisticsTask;
/**
* Interface defines and documents the counters and counter namespaces for
* the {@link ConcurrencyManager}.
*
* @author Bryan Thompson
*/
public static interface IConcurrencyManagerCounters {
/**
* The service to which historical read tasks are submitted.
*/
String ReadService = "Read Service";
/**
* The service to which isolated write tasks are submitted.
*/
String TXWriteService = "Transaction Write Service";
/**
* The service to which {@link ITx#UNISOLATED} tasks are submitted. This
* is the service that handles commit processing. Tasks submitted to
* this service are required to declare resource lock(s) and must
* acquire those locks before they can begin executing.
*/
String writeService = "Unisolated Write Service";
/**
* The performance counters for the object which manages the resource
* locks a {@link WriteExecutorService}. These counters are reported as
* children of the {@link WriteExecutorService}'s counters.
*/
String LockManager = "LockManager";
}
/**
* Reports the elapsed time since the service was started.
*
* @author Bryan Thompson
*/
private static class ServiceElapsedTimeInstrument extends Instrument {
final long serviceStartTime;
public ServiceElapsedTimeInstrument(final long serviceStartTime) {
this.serviceStartTime = serviceStartTime;
}
@Override
public void sample() {
setValue(System.currentTimeMillis() - serviceStartTime);
}
}
/**
* Return the {@link CounterSet}.
*/
@Override
// synchronized
public CounterSet getCounters() {
// if (countersRoot == null){
final CounterSet countersRoot = new CounterSet();
// elapsed time since the service started (milliseconds).
countersRoot.addCounter("elapsed",
new ServiceElapsedTimeInstrument(serviceStartTime));
if (collectQueueStatistics) {
// readService
{
countersRoot.makePath(
IConcurrencyManagerCounters.ReadService).attach(
readServiceQueueStatisticsTask.getCounters());
}
// txWriteService
{
countersRoot.makePath(
IConcurrencyManagerCounters.TXWriteService).attach(
txWriteServiceQueueStatisticsTask.getCounters());
}
// writeService
{
countersRoot.makePath(
IConcurrencyManagerCounters.writeService).attach(
writeServiceQueueStatisticsTask.getCounters());
/*
* The lock manager for the write service.
*/
countersRoot
.makePath(
IConcurrencyManagerCounters.writeService
+ ICounterSet.pathSeparator
+ IConcurrencyManagerCounters.LockManager)
.attach(getWriteService().getLockManager().getCounters());
}
}
// }
return countersRoot;
}
// private CounterSet countersRoot;
/**
* Submit a task (asynchronous). Tasks will execute asynchronously in the
* appropriate thread pool with as much concurrency as possible.
*
* Note: Unisolated write tasks will NOT return before the next group commit
* (exceptions may be thrown if the task fails or the commit fails). The
* purpose of group commits is to provide higher throughput for writes on
* the store by only syncing the data to disk periodically rather than after
* every write. Group commits are scheduled by the {@link #commitService}.
* The trigger conditions for group commits may be configured using
* {@link ConcurrencyManager.Options}. If you are using the store in a
* single threaded context then you may set
* {@link Options#WRITE_SERVICE_CORE_POOL_SIZE} to ONE (1) which has the
* effect of triggering commit immediately after each unisolated write.
* However, note that you can not sync a disk more than ~ 30-40 times per
* second so your throughput in write operations per second will never
* exceed that for a single-threaded application writing on a hard disk.
* (Your mileage can vary if you are writing on a transient store or using a
* durable medium other than disk).
*
* Note: The isolated indices used by a read-write transaction are NOT
* thread-safe. Therefore a partial order is imposed over concurrent tasks
* for the same transaction that seek to read or write on
* the same index(s). Full concurrency is allowed when different
* transactions access the same index(s), but write-write conflicts MAY be
* detected during commit processing.
*
* Note: The following exceptions MAY be wrapped by {@link Future#get()} for
* tasks submitted via this method:
*
* - {@link ValidationError}
* - An unisolated write task was attempting to commit the write set for
* a transaction but validation failed. You may retry the entire
* transaction.
* - {@link InterruptedException}
* - A task was interrupted during execution and before the task had
* completed normally. You MAY retry the task, but note that this exception
* is also generated when tasks are cancelled when the journal is being
* {@link #shutdown()} after the timeout has expired or
* {@link #shutdownNow()}. In either of these cases the task will not be
* accepted by the journal.
*
*
*
*
* @param task
* The task.
*
* @return The {@link Future} that may be used to resolve the outcome of the
* task.
*
* @exception RejectedExecutionException
* if task cannot be scheduled for execution (typically the
* queue has a limited capacity and is full)
* @exception NullPointerException
* if task null
*/
@Override
public FutureTask submit(final AbstractTask task) {
assertOpen();
// Note that time the task was submitted for execution.
task.nanoTime_submitTask = System.nanoTime();
if( task.readOnly ) {
/*
* Reads against historical data do not require concurrency control.
*
* The only distinction between a transaction and an unisolated read
* task is the choice of the historical state from which the task
* will read. A ReadOnly transaction reads from the state of the
* index as of the start time of the transaction. A ReadCommitted
* transaction and an unisolated reader both read from the last
* committed state of the index.
*/
if (log.isInfoEnabled())
log.info("Submitted to the read service: "
+ task.getClass().getName() + ", timestamp="
+ task.timestamp);
return submitWithDynamicLatency(task, readService, countersHR);
} else {
if (task.isReadWriteTx) {
/*
* A task that reads from historical data and writes on isolated
* indices backed by a temporary store. Concurrency control is
* required for the isolated indices on the temporary store, but
* not for the reads against the historical data.
*/
if (log.isInfoEnabled())
log.info("Submitted to the transaction service: "
+ task.getClass().getName() + ", timestamp="
+ task.timestamp);
return submitWithDynamicLatency(task, txWriteService, countersTX);
} else {
/*
* A task that reads from and writes on "live" indices. The live
* indices are NOT thread-safe. Concurrency control provides a
* partial order over the executing tasks such that there is
* never more than one task with access to a given live index.
*/
if (log.isInfoEnabled())
log.info("Submitted to the write service: "
+ task.getClass().getName() + ", timestamp="
+ task.timestamp);
return submitWithDynamicLatency(task, getWriteService(), countersUN);
}
}
}
/**
* Logs a warning if a new task is started when the journal is
* over-extended. This is invoked from the logic which dispatches tasks to
* the readService. This is because the asynchronous overflow tasks
* run on the readService, so that is what is most interesting when the
* journal is over extended.
*
* @param task
* The task.
*/
private void journalOverextended(final AbstractTask task) {
final double overextension = getJournalOverextended();
if (overextension >= 2d) {
/*
* Note: This is being used to diagnose a problem where the live
* journal can continue to grow because asynchronous overflow tasks
* are not being scheduled with enough intelligence.
*
* @todo convert to WARN
*/
log.error("overextended=" + (int)overextension + "x : "
+ task.toString());
}
}
/**
* Return the overextension multiplier for the journal. This is the ratio of
* the bytes written on the journal against its nominal maximum extent. For
* example, an overextension of two means that the journal has reached twice
* is nominal maximum extent. This is zero unless we are running in a
* distributed federation.
*
* @return The overextension multipler.
*/
public double getJournalOverextended() {
// Only for the data service with overflow enabled.
if (!resourceManager.isOverflowEnabled()) {
return 0;
}
// And even then only for the distributed federation
try {
if (!(resourceManager.getFederation() instanceof AbstractDistributedFederation)) {
return 0;
}
} catch (UnsupportedOperationException ex) {
// note: thrown by some unit tests, but not by real services.
return 0;
}
final AbstractJournal journal = resourceManager.getLiveJournal();
return ((double)journal.size()) / journal.getMaximumExtent();
}
/**
* Submit a task to a service, dynamically imposing latency on the caller
* based on the #of tasks already in the queue for that service.
*
* @param task
* The task.
* @param service
* The service.
*
* @return The {@link Future}.
*/
private FutureTask submitWithDynamicLatency(
final AbstractTask task, final ExecutorService service,
final TaskCounters taskCounters) {
/*
* Track the total inter-arrival time.
*/
synchronized (taskCounters.lastArrivalNanoTime) {
final long lastArrivalNanoTime = taskCounters.lastArrivalNanoTime
.get();
final long now = System.nanoTime();
final long delta = now - lastArrivalNanoTime;
// cumulative inter-arrival time.
taskCounters.interArrivalNanoTime.addAndGet(delta);
// update timestamp of the last task arrival.
taskCounters.lastArrivalNanoTime.set(now);
}
taskCounters.taskSubmitCount.incrementAndGet();
/*
* Note: The StoreManager (part of the ResourceManager) has some
* asynchronous startup processing where it scans the existing store
* files or creates the initial store file. This code will await the
* completion of service startup processing before permitting a task to
* be submitted. This causes clients to block until we are ready to
* process their tasks.
*/
if (resourceManager instanceof StoreManager) {
if (!((StoreManager) resourceManager).awaitRunning()) {
throw new RejectedExecutionException(
"StoreManager is not available");
}
}
if(service == readService) {
/*
* Log warnings if new tasks are started when the journal is over
* extended.
*/
journalOverextended( task );
}
if(backoff && service instanceof ThreadPoolExecutor) {
final BlockingQueue queue = ((ThreadPoolExecutor) service)
.getQueue();
if (!(queue instanceof SynchronousQueue)) {
/*
* Note: SynchronousQueue is used when there is no limit on the
* #of workers, e.g., when using
* Executors.newCachedThreadPool(). The SynchronousQueue has a
* ZERO capacity. Therefore the logic to test the remaining
* capacity and inject a delay simply does not work for this
* type of queue.
*/
final int queueRemainingCapacity = queue.remainingCapacity();
final int queueSize = queue.size();
if (queue.size() * 1.10 >= queueRemainingCapacity) {
try {
/*
* Note: Any delay here what so ever causes the #of
* tasks in a commit group to be governed primarily by
* the CORE pool size.
*/
if (BigdataStatics.debug)
System.err.print("z");
Thread.sleep(50/* ms */);
} catch (InterruptedException e) {
throw new RuntimeException(e);
}
}
}
}
final FutureTask ft;
if (service instanceof WriteExecutorService) {
final NonBlockingLockManagerWithNewDesign lockManager = ((WriteExecutorService) service)
.getLockManager();
ft = lockManager.submit(task.getResource(), task);
// writeServiceThreadGuard.guard(ft);
} else {
ft = new FutureTask(task);
service.submit(ft);
}
return ft;
}
// /**
// * Actively running for the {@link WriteExecutorService}.
// *
// * FIXME This is really just the same as those tasks actively executing on
// * the {@link WriteExecutorService}. It does not include tasks that are
// * pending execution (for example, awaiting their locks) and does not include
// * tasks that have executed and are pending group commit.
// *
// * @see HA doLocalAbort()
// * should interrupt NSS requests and AbstractTasks
// */
// private final ThreadGuard writeServiceThreadGuard = new ThreadGuard();
/**
* Cancel any running or queued tasks on the {@link WriteExecutorService}.
*
* @see HA doLocalAbort()
* should interrupt NSS requests and AbstractTasks
*
* FIXME Should also shutdown/restart the {@link #txWriteService}
*/
void abortAllTx() {
// final NonBlockingLockManagerWithNewDesign lockManager = writeService.getLockManager();
//
// // interrupt anything running.
// writeServiceThreadGuard.interruptAll();
}
/**
* When true imposes dynamic latency on arriving tasks in
* {@link #submitWithDynamicLatency(AbstractTask, ExecutorService, TaskCounters)}.
*
* @todo revisit the question of imposed latency here based on performance
* analysis (of queue length vs response time) for the federation
* under a variety of workloads (tasks such as rdf data load, rdf data
* query, bigdata repository workloads, etc.).
*
* Note that {@link Executors#newCachedThreadPool()} uses a
* {@link SynchronousQueue} and that queue has ZERO capacity.
*/
static private final boolean backoff = false;
/**
* Executes the given tasks, returning a list of Futures holding their
* status and results when all complete. Note that a completed task could
* have terminated either normally or by throwing an exception. The results
* of this method are undefined if the given collection is modified while
* this operation is in progress.
*
* Note: Contract is per {@link ExecutorService#invokeAll(Collection)}
*
* @param tasks
* The tasks.
*
* @return Their {@link Future}s.
*
* @exception InterruptedException
* if interrupted while waiting, in which case unfinished
* tasks are canceled.
* @exception NullPointerException
* if tasks or any of its elements are null
* @exception RejectedExecutionException
* if any task cannot be scheduled for execution
*/
@Override
public List> invokeAll(
final Collection> tasks)
throws InterruptedException {
assertOpen();
final List> futures = new LinkedList>();
boolean done = false;
try {
// submit all.
for (AbstractTask task : tasks) {
futures.add(submit(task));
}
// await all futures.
for (Future f : futures) {
if (!f.isDone()) {
try {
f.get();
} catch (ExecutionException ex) {
// ignore.
} catch (CancellationException ex) {
// ignore.
}
}
}
done = true;
return futures;
} finally {
if (!done) {
// At least one future did not complete.
for (Future f : futures) {
if(!f.isDone()) {
f.cancel(true/* mayInterruptIfRunning */);
}
}
}
}
}
/**
* Executes the given tasks, returning a list of Futures holding their
* status and results when all complete or the timeout expires, whichever
* happens first. Note that a completed task could have terminated either
* normally or by throwing an exception. The results of this method are
* undefined if the given collection is modified while this operation is in
* progress.
*
* Note: Contract is based on
* {@link ExecutorService#invokeAll(Collection, long, TimeUnit)} but only
* the {@link Future}s of the submitted tasks are returned.
*
* @param tasks
* The tasks.
*
* @return The {@link Future}s of all tasks that were
* {@link #submit(AbstractTask) submitted} prior to the expiration
* of the timeout.
*
* @exception InterruptedException
* if interrupted while waiting, in which case unfinished
* tasks are canceled.
* @exception NullPointerException
* if tasks or any of its elements are null
* @exception RejectedExecutionException
* if any task cannot be scheduled for execution
*/
@Override
public List> invokeAll(
final Collection> tasks, final long timeout,
final TimeUnit unit) throws InterruptedException {
assertOpen();
final List> futures = new LinkedList>();
boolean done = false;
long nanos = unit.toNanos(timeout);
long lastTime = System.nanoTime();
try {
// submit all.
for (AbstractTask task : tasks) {
final long now = System.nanoTime();
nanos -= now - lastTime;
lastTime = now;
if (nanos <= 0) {
// timeout.
return futures;
}
futures.add(submit(task));
}
// await all futures.
for (Future f : futures) {
if (!f.isDone()) {
if (nanos <= 0) {
// timeout
return futures;
}
try {
f.get(nanos, TimeUnit.NANOSECONDS);
} catch (TimeoutException ex) {
if (log.isInfoEnabled()) log.info("Task Timeout");
return futures;
} catch (ExecutionException ex) {
// ignore.
} catch (CancellationException ex) {
// ignore.
}
final long now = System.nanoTime();
nanos -= now - lastTime;
lastTime = now;
}
}
done = true;
return futures;
} finally {
if (!done) {
// At least one future did not complete.
for (Future f : futures) {
if (!f.isDone()) {
f.cancel(true/* mayInterruptIfRunning */);
}
}
}
}
}
}