io.github.udaysagar2177:aurora-maps

aurora-maps

The goal of this repository is to create fastest off-heap maps in Java for non-concurrent use cases.

Currently, Chronicle maps are the best choice to store map data off-heap but due to its concurrency support, it delivers poor performance for non-concurrent use cases. Performance target of this repository is to get close to efficient on-heap implementations like Koloboke maps. To keep it simple, only fixed size key-values are supported. If you need support for variable size key-values, extra effort must be taken to store the data of variable size at some other location and use the indices to that data in these maps.

Quick start:

Add the following dependency in your Maven pom.xml:

<dependency>
  <groupId>io.github.udaysagar2177</groupId>
  <artifactId>aurora-maps</artifactId>
  <version>0.0.1</version>
</dependency>

Then, you can initialize and use a new off-heap map as shown below.

Example:

public class IntIntMapExample {
    /**
     * Simple example that shows how to build a int-int map on off-heap memory.
     */
    public void example() throws Exception {

        // map creation
        OffHeapMapImpl<IntHolder, IntHolder> map = new OffHeapMapImpl<>(1, 0.66f,
                new IntIntEntrySeDeserializer(), DirectMemoryResource::new,
                IntHolder::new, IntHolder::new);

        try {
            // flyweights to be used to put or get data
            IntHolder key = new IntHolder();
            IntHolder putValue = new IntHolder();
            IntHolder getValue = new IntHolder();

            // put
            map.put(key.setInt(12), putValue.setInt(20), getValue);
            // get
            map.get(key, getValue);
            // remove
            map.remove(key, getValue);
            // clear for re-use
            map.clear();
        } finally {
            // close map and release off-heap memory
            map.close();
        }
    }
}

This repository also provides support for short-short, long-long, byte[]-byte[] key values through ShortShortEntrySeDeserializer, LongLongEntrySeDeserializer and ByteArrByteArrEntrySeDeserializer. Additional data types support can be added by extending EntrySeDeserializer interface.

Mmap data to disk

If you want to create the map with memory mapped to disk, you can simply replace DirectMemoryResource::new in the above example with (capacityInBytes) -> new MmapMemoryResource("/tmp", "intIntMap", capacityInBytes). Here, "/tmp" is the folder name inside which the data will be mapped. "intIntMap" is the file prefix used to create new files when scaling up the map. Same folder/file name can be used for multiple map objects because MmapMemoryResource appends an increasing id to all new file names.

Performance comparison:

primitive int-int map

Single get performance

Benchmark                                   Mode   Cnt         Score         Error  Units
IntIntMapBenchmark.testGetOnChronicleMap    thrpt  100   9749814.908 ±  226087.135  ops/s
IntIntMapBenchmark.testGetOnKolobokeMap     thrpt  100  24617367.499 ±  878137.939  ops/s
IntIntMapBenchmark.testGetOnMmapOffHeapMap  thrpt  100  20850211.560 ± 1047023.575  ops/s
IntIntMapBenchmark.testGetOnOffHeapMap      thrpt  100  19598000.585 ±  714347.341  ops/s

1 million insertions:

Benchmark                                    Mode  Cnt         Score         Error  Units
IntIntMapBenchmark.testPutOnChronicleMap    thrpt  100         1.633 ±       0.035  ops/s
IntIntMapBenchmark.testPutOnKolobokeMap     thrpt  100        24.604 ±       0.895  ops/s
IntIntMapBenchmark.testPutOnMmapOffHeapMap  thrpt  100        15.996 ±       0.601  ops/s
IntIntMapBenchmark.testPutOnOffHeapMap      thrpt  100        16.791 ±       0.618  ops/s

byte[]-byte[] map

• key size: 24, value size: 42

Single get performance

Benchmark                                         Mode  Cnt        Score      Error  Units
ByteArrByteArrBenchmark.testGetOnChronicleMap    thrpt   60  2339801.966 ± 7619.689  ops/s
ByteArrByteArrBenchmark.testGetOnKolobokeMap     thrpt   60  3730351.924 ± 3920.856  ops/s
ByteArrByteArrBenchmark.testGetOnMmapOffHeapMap  thrpt   60  2499914.883 ± 5112.507  ops/s
ByteArrByteArrBenchmark.testGetOnOffHeapMap      thrpt   60  2496837.620 ± 3866.586  ops/s

1 million insertions:

Benchmark                                         Mode  Cnt        Score      Error  Units
ByteArrByteArrBenchmark.testPutOnChronicleMap    thrpt   60        0.927 ±    0.005  ops/s
ByteArrByteArrBenchmark.testPutOnKolobokeMap     thrpt   60        1.326 ±    0.002  ops/s
ByteArrByteArrBenchmark.testPutOnMmapOffHeapMap  thrpt   60        0.856 ±    0.017  ops/s
ByteArrByteArrBenchmark.testPutOnOffHeapMap      thrpt   60        0.984 ±    0.001  ops/s

Integer key-value maps are 10x faster than ChronicleMap on puts and 2x on gets. But byte array key-value maps are performing same as ChronicleMap, that needs to be investigated and improved.

TODO:

1. More performance improvements
2. Add append-only map wrapper, i.e., a map that doesn't support removal and key-value overrides for single writer multiple readers use case
3. Move chronicle map dependency to benchmarks folder

Thanks to https://github.com/mikvor/hashmapTest & https://github.com/OpenHFT/Chronicle-Map for inspiring this work.