实现不同步读取的双缓冲Java HashMap

Question

So I thought I had this genius idea to solve a pretty specific problem, but I can't get rid of one last potential thread safety problem. I was wondering if you guys would have an idea to solve it.

The problem:

A huge number of threads need to read from a HashMap that only rarely updates. The problem is that in ConcurrentHashMap, ie, the thread safe version, the read methods still have the potential to hit a mutex, as write methods still lock bins (ie, sections of the map).

The idea:

Have 2 hidden HashMaps acting as one... one for threads to read with no synchronization, the other for threads to write in, with synchronization of course, and every once in a while, flip them.

The obvious caveat is that the map is only eventually consistent, but let's assume that this is good enough for it's intended purpose.

But the problem that's come up is that it still leaves one race condition open, even when using AtomicInteger and such, because just when the flip happens, I can't be sure a reader didn't slip in... The problem is between line 262-272 in the startRead() method and line 241-242 in the flip() method.

---

Obviously ConcurrentHashMap is a very very good class to use for this problem, I just want to see if I can push the idea a little further.

Anyone have any ideas?

---

Here's the full code of the class. (Not fully debugged/tested, but you get the idea...)

    package org.nectarframework.base.tools;

    import java.util.Collection;

    import java.util.HashMap;
    import java.util.LinkedList;
    import java.util.Map;
    import java.util.Set;
    import java.util.concurrent.atomic.AtomicBoolean;
    import java.util.concurrent.atomic.AtomicInteger;

    /**
     * 
     * This map is intended to be both thread safe, and have (mostly) non mutex'd
     * reads.
     * 
     * HOWEVER, if you insert something into this map, and immediately try to read
     * the same key from the map, it probably won't give you the result you expect.
     * 
     * The idea is that this map is in fact 2 maps, one that handles writes, the
     * other reads, and every so often the two maps switch places.
     * 
     * As a result, this map will be eventually consistent, and while writes are
     * still synchronized, reads are not.
     * 
     * This map can be very effective if handling a massive number of reads per unit
     * time vs a small number of writes per unit time, especially in a massively
     * multithreaded use case.
     * 
     * This class isn't such a good idea because it's possible that between
     * readAllowed.get() and readCounter.increment(), the flip() happens,
     * potentially sending one or more threads on the Map that flip() is about to
     * update. The solution would be an
     * AtomicInteger.compareGreaterThanAndIncrement(), but that doesn't exist.
     * 
     * 
     * @author schuttek
     *
     */

    public class DoubleBufferHashMap<K, V> implements Map<K, V> {

        private Map<K, V> readMap = new HashMap<>();
        private Map<K, V> writeMap = new HashMap<>();
        private LinkedList<Triple<Operation, Object, V>> operationList = new LinkedList<>();

        private AtomicBoolean readAllowed = new AtomicBoolean(true);
        private AtomicInteger readCounter = new AtomicInteger(0);

        private long lastFlipTime = System.currentTimeMillis();
        private long flipTimer = 3000; // 3 seconds

        private enum Operation {
            Put, Delete;
        }

        @Override
        public int size() {
            startRead();
            RuntimeException rethrow = null;
            int n = 0;
            try {
                n = readMap.size();
            } catch (RuntimeException t) {
                rethrow = t;
            }
            endRead();
            if (rethrow != null) {
                throw rethrow;
            }
            return n;
        }

        @Override
        public boolean isEmpty() {
            startRead();
            RuntimeException rethrow = null;
            boolean b = false;
            try {
                b = readMap.isEmpty();
            } catch (RuntimeException t) {
                rethrow = t;
            }
            endRead();
            if (rethrow != null) {
                throw rethrow;
            }
            return b;
        }

        @Override
        public boolean containsKey(Object key) {
            startRead();
            RuntimeException rethrow = null;
            boolean b = false;
            try {
                b = readMap.containsKey(key);
            } catch (RuntimeException t) {
                rethrow = t;
            }
            endRead();
            if (rethrow != null) {
                throw rethrow;
            }
            return b;
        }

        @Override
        public boolean containsValue(Object value) {
            startRead();
            RuntimeException rethrow = null;
            boolean b = false;
            try {
                b = readMap.containsValue(value);
            } catch (RuntimeException t) {
                rethrow = t;
            }
            endRead();
            if (rethrow != null) {
                throw rethrow;
            }
            return b;
        }

        @Override
        public V get(Object key) {
            startRead();
            RuntimeException rethrow = null;
            V v = null;
            try {
                v = readMap.get(key);
            } catch (RuntimeException t) {
                rethrow = t;
            }
            endRead();
            if (rethrow != null) {
                throw rethrow;
            }
            return v;
        }

        @Override
        public synchronized V put(K key, V value) {
            operationList.add(new Triple<>(Operation.Put, key, value));
            writeMap.put(key, value);
            return value;
        }

        @Override
        public synchronized V remove(Object key) {
            // Not entirely sure if we should return the value from the read map or
            // the write map...
            operationList.add(new Triple<>(Operation.Delete, key, null));
            V v = writeMap.remove(key);
            endRead();
            return v;
        }

        @Override
        public synchronized void putAll(Map<? extends K, ? extends V> m) {
            for (K k : m.keySet()) {
                V v = m.get(k);
                operationList.add(new Triple<>(Operation.Put, k, v));
                writeMap.put(k, v);
            }
            checkFlipTimer();
        }

        @Override
        public synchronized void clear() {
            writeMap.clear();
            checkFlipTimer();
        }

        @Override
        public Set<K> keySet() {
            startRead();
            RuntimeException rethrow = null;
            Set<K> sk = null;
            try {
                sk = readMap.keySet();
            } catch (RuntimeException t) {
                rethrow = t;
            }
            endRead();
            if (rethrow != null) {
                throw rethrow;
            }
            return sk;
        }

        @Override
        public Collection<V> values() {
            startRead();
            RuntimeException rethrow = null;
            Collection<V> cv = null;
            try {
                cv = readMap.values();
            } catch (RuntimeException t) {
                rethrow = t;
            }
            endRead();
            if (rethrow != null) {
                throw rethrow;
            }
            return cv;
        }

        @Override
        public Set<java.util.Map.Entry<K, V>> entrySet() {
            startRead();
            RuntimeException rethrow = null;
            Set<java.util.Map.Entry<K, V>> se = null;
            try {
                se = readMap.entrySet();
            } catch (RuntimeException t) {
                rethrow = t;
            }
            endRead();
            if (rethrow != null) {
                throw rethrow;
            }
            endRead();
            return se;
        }

        private void checkFlipTimer() {
            long now = System.currentTimeMillis();
            if (this.flipTimer > 0 && now > this.lastFlipTime + this.flipTimer) {
                flip();
                this.lastFlipTime = now;
            }
        }

        /**
         * Flips the two maps, and updates the map that was being read from to the
         * latest state.
         */
        @SuppressWarnings("unchecked")
        private synchronized void flip() {
            readAllowed.set(false);
            while (readCounter.get() != 0) {
                Thread.yield();
            }

            Map<K, V> temp = readMap;
            readMap = writeMap;
            writeMap = temp;

            readAllowed.set(true);
            this.notifyAll();

            for (Triple<Operation, Object, V> t : operationList) {
                switch (t.getLeft()) {
                case Delete:
                    writeMap.remove(t.getMiddle());
                    break;
                case Put:
                    writeMap.put((K) t.getMiddle(), t.getRight());
                    break;
                }
            }
        }

        private void startRead() {
            if (!readAllowed.get()) {
                synchronized (this) {
                    try {
                        wait();
                    } catch (InterruptedException e) {
                    }
                }
            }
            readCounter.incrementAndGet();
        }

        private void endRead() {
            readCounter.decrementAndGet();
        }

    }

Answer 1

I strongly suggest you to learn how to use JMH , which is the first thing you should learn on the path of optimizing algorithms and data-structures.

For example if you know how to use it, you can quickly find that when there is only 10% of writes ConcurrentHashMap performs very close to unsynchronized HashMap .

4 Threads (10% writes):

Benchmark                      Mode  Cnt   Score   Error  Units
SO_Benchmark.concurrentMap    thrpt    2  69,275          ops/s
SO_Benchmark.usualMap         thrpt    2  78,490          ops/s

8 Threads (10% writes):

Benchmark                      Mode  Cnt    Score   Error  Units
SO_Benchmark.concurrentMap    thrpt    2   93,721          ops/s
SO_Benchmark.usualMap         thrpt    2  100,725          ops/s

With smaller percentage of writes ConcurrentHashMap 's performance tends to go even more close to HashMap 's one.

Now I modified your startRead and endRead , and made them non-functional, but very simple:

private void startRead() {
    readCounter.incrementAndGet();
    readAllowed.compareAndSet(false, true);
}

private void endRead() {
    readCounter.decrementAndGet();
    readAllowed.compareAndSet(true, false);
}

And lets look at the performance:

Benchmark                      Mode  Cnt    Score   Error  Units
SO_Benchmark.concurrentMap    thrpt   10   98,275 ? 2,018  ops/s
SO_Benchmark.doubleBufferMap  thrpt   10   80,224 ? 8,993  ops/s
SO_Benchmark.usualMap         thrpt   10  106,224 ? 4,205  ops/s

These results show us that with one atomic counter and one atomic boolean modification on each operation we can't get better performance than ConcurrentHashMap . (I've tried 30,10 and 5 percentage of writes, but it never resulted in better performance with DoubleBufferHashMap )

Pastebin with benchmark if you are interested.