...
The Java maps are known to be non-storage friendly. It's assumed that often 80%+ of the map size is used for maintenance, where only 20% are real data stored in the map. Note that we use ConcurrentHashMap.
...
Current size
A test was performed to check the amount of indexed elements in the maps as well as the size of the complete indexing tree. The test was executed with normal load pace and invocations having ~50-100 children. Max size of buffer is 335MB. The results are:
| Code Block | ||
|---|---|---|
| ||
2014-11-14 15:44:18,125: 1239302 [indexing-thread] DEBUG it.cmr.cache.impl.AtomicBuffer - Indexing tree new size: 32934038
2014-11-14 15:44:18,125: 1239302 [indexing-thread] DEBUG it.cmr.cache.impl.AtomicBuffer - Indexing occupancy percentage: 10.049% |
| Code Block | ||
|---|---|---|
| ||
2014-11-14 15:49:13,842: 1535019 [pool-2-thread-1] INFO ctit.indexing.buffer.impl.Leaf - Leaf map size prior to clean = 17245
2014-11-14 15:49:13,843: 1535020 [pool-2-thread-1] INFO ctit.indexing.buffer.impl.Leaf - Leaf map size prior to clean = 300778
2014-11-14 15:49:13,843: 1535020 [pool-2-thread-1] INFO ctit.indexing.buffer.impl.Leaf - Leaf map size prior to clean = 4554
2014-11-14 15:49:13,843: 1535020 [pool-2-thread-1] INFO ctit.indexing.buffer.impl.Leaf - Leaf map size prior to clean = 32965
2014-11-14 15:49:13,843: 1535020 [pool-2-thread-1] INFO ctit.indexing.buffer.impl.Leaf - Leaf map size prior to clean = 644
2014-11-14 15:49:13,843: 1535020 [pool-2-thread-1] INFO ctit.indexing.buffer.impl.Leaf - Leaf map size prior to clean = 10604
2014-11-14 15:49:13,843: 1535020 [pool-2-thread-1] INFO ctit.indexing.buffer.impl.Leaf - Leaf map size prior to clean = 598
2014-11-14 15:49:13,843: 1535020 [pool-2-thread-1] INFO ctit.indexing.buffer.impl.Leaf - Leaf map size prior to clean = 10604 |
As seen, indexing tree in this situation occupies 10% of buffer, totaling to over 30MB. In my opinion this is ideal situation and under 10% we can not go.
We have total of 8 maps, with different sizes, biggest one having 300,778 indexed elements.
ConcurrentHashMap compared with other thread-safe map
We furthermore tested the ConcurrentHashMap with two other map implementations: NonBlockingHashMapLong from Cliff's Click high scale lib and synchronized Long2ObjectMap from FastUtil lib. The results shows that NonBlockingHashMap beats heavily other maps in both size and performance measurements:
Size
| Number of elements | ConcurrentHashMap | NonBlockingHashMapLong | Long2ObjectMap | |||
|---|---|---|---|---|---|---|
| Size | Ch. (%) | Size | Ch. (%) | Size | Ch. (%) | |
| Empty | 224 | - | 712 | +317% | 568 | +253% |
| 8 | 800 | - | 824 | +3% | 696 | -13% |
| 32 | 2720 | - | 1400 | -48% | 1496 | -45% |
| 256 | 20640 | - | 7672 | -63% | 10904 | -47% |
| 1024 | 82080 | - | 29176 | -64% | 43160 | -47% |
| 4096 | 327840 | - | 115192 | -65% | 172184 | -47% |
As seen in the table there are three stages of elements numbers:
- empty map makes Java implementation the smallest
- with 8 elements have maps similar size
- from 256 elements the max saving for NonBlickingHashMapLong and Long2ObjectMap is achieved
Performance
| Code Block | ||
|---|---|---|
| ||
Java
_______________________
0 - 2 producers 2 consumers: 4,782,614 ops/sec - MapsPerfTest
1 - 2 producers 2 consumers: 4,194,545 ops/sec - MapsPerfTest
2 - 2 producers 2 consumers: 4,726,670 ops/sec - MapsPerfTest
3 - 2 producers 2 consumers: 5,339,190 ops/sec - MapsPerfTest
4 - 2 producers 2 consumers: 4,952,414 ops/sec - MapsPerfTest
Cliff's
______________________
0 - 2 producers 2 consumers: 23,030,573 ops/sec - MapsPerfTest
1 - 2 producers 2 consumers: 31,244,208 ops/sec - MapsPerfTest
2 - 2 producers 2 consumers: 30,630,539 ops/sec - MapsPerfTest
3 - 2 producers 2 consumers: 28,825,699 ops/sec - MapsPerfTest
4 - 2 producers 2 consumers: 26,588,300 ops/sec - MapsPerfTest
Long2ObjectMap (FastUtil)
______________________
0 - 2 producers 2 consumers: 5,569,578 ops/sec - MapsPerfTest
1 - 2 producers 2 consumers: 5,940,526 ops/sec - MapsPerfTest
2 - 2 producers 2 consumers: 5,883,532 ops/sec - MapsPerfTest
3 - 2 producers 2 consumers: 5,798,396 ops/sec - MapsPerfTest
4 - 2 producers 2 consumers: 5,537,185 ops/sec - MapsPerfTest |