BerkeleyDB JE random access time increases non-linearly

Question

I am testing BerkeleyDB Java Edition to understand whether I can use it in my project.

I've created very simple program which works with object of class com.sleepycat.je.Database:

• writes N records of 5-15kb each, with keys generated like Integer.toString(random.nextInt());

• reads these records fetching them with method Database#get in the same order they were created;

• reads the same number of records with method Database#get in random order.

And I now see the strange thing. Execution time for third test grows very non-linearly with increasing of the number of records.

• N=80000, write=55sec, sequential fetch=17sec, random fetch=3sec
• N=100000, write=60sec, sequential fetch=20sec, random fetch=7sec
• N=120000, write=68sec, sequential fetch=27sec, random fetch=11sec
• N=140000, write=82sec, sequential fetch=32sec, random fetch=47sec

(I've run tests several times, of course.)

I suppose I am doing something quite wrong. Here is the source for reference (sorry, it is bit long), methods are called in the same order:

private Environment env;
private Database db;
private Random random = new Random();
private List<String> keys = new ArrayList<String>();
private int seed = 113;


public boolean dbOpen() {
    EnvironmentConfig ec = new EnvironmentConfig();
    DatabaseConfig dc = new DatabaseConfig();
    ec.setAllowCreate(true);
    dc.setAllowCreate(true);
    env = new Environment(new File("mydbenv"), ec);
    db = env.openDatabase(null, "moe", dc);
    return true;
}

public int storeRecords(int i) {
    int j;
    long size = 0;
    DatabaseEntry key = new DatabaseEntry();
    DatabaseEntry val = new DatabaseEntry();

    random.setSeed(seed);

    for (j = 0; j < i; j++) {
        String k = Long.toString(random.nextLong());
        byte[] data = new byte[5000 + random.nextInt(10000)];
        keys.add(k);

        size += data.length;

        random.nextBytes(data);
        key.setData(k.getBytes());
        val.setData(data);
        db.put(null, key, val);
    }

    System.out.println("GENERATED SIZE: " + size);

    return j;
}                   

public int fetchRecords(int i) {
    int j, res;
    DatabaseEntry key = new DatabaseEntry();
    DatabaseEntry val = new DatabaseEntry();

    random.setSeed(seed);
    res = 0;

    for (j = 0; j < i; j++) {
        String k = Long.toString(random.nextLong());
        byte[] data = new byte[5000 + random.nextInt(10000)];
        random.nextBytes(data);
        key.setData(k.getBytes());
        db.get(null, key, val, null);
        if (Arrays.equals(data, val.getData())) {
            res++;
        } else {
            System.err.println("FETCH differs: " + j);
            System.err.println(data.length + " " + val.getData().length);
        }
    }

    return res;
}

public int fetchRandom(int i) {
    DatabaseEntry key = new DatabaseEntry();
    DatabaseEntry val = new DatabaseEntry();

    for (int j = 0; j < i; j++) {
        String k = keys.get(random.nextInt(keys.size()));
        key.setData(k.getBytes());
        db.get(null, key, val, null);
    }

    return i;
}

Answer 1

Performance degradation is non-linear for two reasons:

1. BDB-JE data structure is a b-tree, which has O(log(n)) performance for retrieving one record. Retrieving all via the get method is O(n*log(n)).
2. Large data sets don't fit into RAM, and so disk access slows everything down. Random access has very poor cache locality.

Note that you can improve write performance by giving up some durability: ec.setTxnWriteNoSync(true);

You might also want to try Tupl, an open source BerkeleyDB replacement I've been working on. It's still in the alpha stage, but you can find it on SourceForge.

For a fair comparison between BDB-JE and Tupl, I set the cache size to 500M and an explicit checkpoint is performed at the end of the store method.

With BDB-JE:

• N=80000, write=11.0sec, fetch=5.3sec
• N=100000, write=13.6sec, fetch=7.0sec
• N=120000, write=16.4sec, fetch=29.5sec
• N=140000, write=18.8sec, fetch=35.9sec
• N=160000, write=21.5sec, fetch=41.3sec
• N=180000, write=23.9sec, fetch=46.4sec

With Tupl:

• N=80000, write=21.7sec, fetch=4.4sec
• N=100000, write=27.6sec, fetch=6.3sec
• N=120000, write=30.2sec, fetch=8.4sec
• N=140000, write=35.4sec, fetch=12.2sec
• N=160000, write=39.9sec, fetch=17.4sec
• N=180000, write=45.4sec, fetch=22.8sec

BDB-JE is faster at writing entries, because of its log-based format. Tupl is faster at reading, however. Here's the source to the Tupl test:

import java.io.; import java.util.;

import org.cojen.tupl.*;

public class TuplTest { public static void main(final String[] args) throws Exception { final RandTupl rt = new RandTupl(); rt.dbOpen(args[0]);

    {
        long start = System.currentTimeMillis();
        rt.storeRecords(Integer.parseInt(args[1]));
        long end = System.currentTimeMillis();
        System.out.println("store duration: " + (end - start));
    }

    {
        long start = System.currentTimeMillis();
        rt.fetchRecords(Integer.parseInt(args[1]));
        long end = System.currentTimeMillis();
        System.out.println("fetch duration: " + (end - start));
    }
}

private Database db;
private Index ix;
private Random random = new Random();
private List<String> keys = new ArrayList<String>();
private int seed = 113;

public boolean dbOpen(String home) throws Exception {
    DatabaseConfig config = new DatabaseConfig();
    config.baseFile(new File(home));
    config.durabilityMode(DurabilityMode.NO_FLUSH);
    config.minCacheSize(500000000);
    db = Database.open(config);
    ix = db.openIndex("moe");
    return true;
}

public int storeRecords(int i) throws Exception {
    int j;
    long size = 0;

    random.setSeed(seed);

    for (j = 0; j < i; j++) {
        String k = Long.toString(random.nextLong());
        byte[] data = new byte[5000 + random.nextInt(10000)];
        keys.add(k);

        size += data.length;

        random.nextBytes(data);
        ix.store(null, k.getBytes(), data);
    }

    System.out.println("GENERATED SIZE: " + size);

    db.checkpoint();
    return j;
}

public int fetchRecords(int i) throws Exception {
    int j, res;

    random.setSeed(seed);
    res = 0;

    for (j = 0; j < i; j++) {
        String k = Long.toString(random.nextLong());
        byte[] data = new byte[5000 + random.nextInt(10000)];
        random.nextBytes(data);
        byte[] val = ix.load(null, k.getBytes());
        if (Arrays.equals(data, val)) {
            res++;
        } else {
            System.err.println("FETCH differs: " + j);
            System.err.println(data.length + " " + val.length);
        }
    }

    return res;
}

public int fetchRandom(int i) throws Exception {
    for (int j = 0; j < i; j++) {
        String k = keys.get(random.nextInt(keys.size()));
        ix.load(null, k.getBytes());
    }

    return i;
}

}