package test;
import java.util.Collections; import java.util.HashMap; import java.util.HashSet; import java.util.Map; import java.util.Set; import java.util.concurrent.ConcurrentHashMap; import java.util.concurrent.Executors; import java.util.concurrent.ScheduledExecutorService; import java.util.concurrent.ThreadFactory; import java.util.concurrent.TimeUnit; import java.util.concurrent.locks.ReentrantReadWriteLock;
/** * Thread safe object map, which clears objects after a specified time. The objects are stored in the * underlying hashMap. * @author rongkan * */ public final class TimeLimitedCacheMap { private final ConcurrentHashMap<String, Object> objectMap = new ConcurrentHashMap<String, Object>(10); private final ConcurrentHashMap<String, Long> timeMap = new ConcurrentHashMap<String, Long>(); /* I need a shared lock, readwrite lock is an excellent candidate. * evcition is run with writeLock, put/remove with readLock */ private final ReentrantReadWriteLock accessLock = new ReentrantReadWriteLock(); //private final ReentrantLock accessLock = new ReentrantLock(); private final Runnable evictor = new Runnable() { /* evictor thread removes data, and changes map state. This is * in conflict with put() and remove(). So we need sync for these operations * * In case you are wondering why evictor needs sync (it just calls remove() right?) * eviction is a compound action that spans more than a single remove(). It enters * into a conflict with put() * * evictor: start looping ------------------------------> keyset is stale, evictor removes the recent put & armagedon comes * Thrd1:--------------------->put(same key, new object) * */ @Override public void run() { // avoid runs on empty maps if(timeMap.isEmpty()){ Thread.yield(); } long currentTime = System.nanoTime(); accessLock.writeLock().lock(); Set<String> keys = new HashSet<String>(timeMap.keySet()); accessLock.writeLock().unlock(); /* First attempt to detect & mark stale entries, but don't delete * The hash map may contain 1000s of objects dont' block it. The returned * Set returned may be stale, implying: * 1. contains keys for objects which are removed by user, using remove() (not a problem) * 2. contains keys for objects which are updated by user, using put() [a big problem] */ Set<String> markedForRemoval = new HashSet<String>(10); for (String key : keys) { long lastTime = timeMap.get(key); if(lastTime == 0){ continue; } long interval = currentTime - lastTime; long elapsedTime = TimeUnit.NANOSECONDS.convert(interval, expiryTimeUnit); if(elapsedTime > expiryTime){ markedForRemoval.add(key); } } /* Actual removal call, which runs on the objects marked earlier. * Assumption: marked objects.size() < hashmap.size() * Do not delete blindly, check for staleness before calling remove */ accessLock.writeLock().lock(); for (String key : markedForRemoval) { long lastTime = timeMap.get(key); if(lastTime == 0){ continue; } long interval = currentTime - lastTime; long elapsedTime = TimeUnit.NANOSECONDS.convert(interval, expiryTimeUnit); if(elapsedTime > expiryTime){ remove(key); } } accessLock.writeLock().unlock(); } }; private final ScheduledExecutorService timer = Executors.newSingleThreadScheduledExecutor(new MyThreadFactory(true)); private final class MyThreadFactory implements ThreadFactory { private boolean isDaemon = false; public MyThreadFactory(boolean daemon){ isDaemon = daemon; }
@Override public Thread newThread(Runnable r) { Thread t = new Thread(r); t.setDaemon(isDaemon); return t; } }; private final long expiryTime; private final TimeUnit expiryTimeUnit; /** * Users can play around with evictionDelay and expiryTime. * 1. Large evictionDelay => less frequent checks, hence chances of finding expired Objects are more * 2. Lean evictionDelay => aggressive checks, and hence more sync overhead with put() and remove() * 3. Large expiryTime => increases the time object stays in object map and less chance of cache miss (cache miss is bad) * 4. Lean expiryTime => by itself does not force object to be removed aggressively, needs lean eviction to be configured * * In case you are wondering, above picture is not complete. * Another key aspect is "Arrival Periodicty", or the rate at which put() is called. * * Ideally: expiryTime == arrival periodicity + 1 [read '+1' as slightly greater] * evictionDelay == expiryTime + 1 * * For random arrival times, which is a more common scenario, use the following pointers * 1. eviction Delay > expiry Time * Here user needs to think of the impact of stale hit (define how stale is stale!) * 2. eviction Delay < arrival time * This has higher chances of cache miss and accidental treatment as failure * * 3. eviction Delay < expiry Time * Unwanted eviction run(s) resulting in sync overhead on map * 4. eviction Delay > arrival Time * Unwanted eviction run(s) resulting in sync overhead on map * * @param initialDelay, time after which scheduler starts * @param evictionDelay, periodicity with which eviction is carried out * @param expiryTime, age of the object, exceeding which the object is to be removed * @param unit */ public TimeLimitedCacheMap(long initialDelay, long evictionDelay, long expiryTime, TimeUnit unit){ timer.scheduleWithFixedDelay(evictor, initialDelay, evictionDelay, unit); this.expiryTime = expiryTime; this.expiryTimeUnit = unit; }
/* The intention is to prevent user from modifying the object Map, * I want all adds/removals to be suspended. synchronizing on objectMap * would also work, but locks are easier to read without scope issues of {} * * Concurrent hashMap would have allowed put and remove to happen concurrently. * I did not use conc-map, because * 1. I want to update another map (timeMap) * 2. I want to sync the operation with evictor thread * * The unfortunate invariants: * 1. sync between evictor and put() * 2. sync between evictor and remove() * imply * 3. sync lock between put() and remove() * * 3. is unfortunate side effect, as you need to sync all exposed invariants on the same lock. * Lock duplication won't help. If I create putLock() and removeLock(), they will allow put() and remove() * to happen concurrently, but will not help two put()/remove() calls to happen in parallel. */ public void put(String key, Object value) { accessLock.readLock().lock(); Long nanoTime = System.nanoTime(); timeMap.put(key, nanoTime); objectMap.put(key, value); accessLock.readLock().unlock(); }
/* Read comments for put() they apply here as well. * If had not allowed remove(), life would have been zillion times simpler. * However, an undoable action is quite bad. */ public Object remove(Object key) { accessLock.readLock().lock(); //accessLock.lock(); Object value = objectMap.remove(key); timeMap.remove(key); //accessLock.unlock(); accessLock.readLock().unlock(); return value; } /* Clone requires locking, to prevent the edge case where * the map is updated with clone is in progress */ public Map<String, Object> getClonedMap(){ accessLock.writeLock().lock(); HashMap<String, Object> mapClone = new HashMap<String, Object>(objectMap); accessLock.writeLock().unlock(); return Collections.unmodifiableMap(mapClone); }
}
|
2 comments:
Great stuff! Just what I needed thank you!
Very nice Article
Post a Comment