jhrotko commented on PR #949:
URL: https://github.com/apache/arrow-java/pull/949#issuecomment-3729300767
Test Configuration
- Vector Length w/ 10,000 elements
- Old Version: Commit e35ed7c3 (UuidHolder with `ArrowBuf buffer` + `int
start`)
- New Version: uuid-improvement branch (UuidHolder with `long mostSigBits` +
`long leastSigBits`)
- JMH Settings: 2 warmup iterations, 3 measurement iterations, GC profiler
enabled
## 1. getUuidDirectly
Read UUID objects directly via `vector.getObject(i)`
| Metric | Old (buffer+offset) | New (two-longs) | Change |
|--------|---------------------|-----------------|--------|
| **Time (µs/op)** | 10.007 ± 1.404 | 9.960 ± 0.460 | **-0.5%** (0.047 µs
faster) |
| **Memory Allocation (B/op)** | 0.012 ± 0.147 | 0.012 ± 0.145 | **Same** |
| **Allocation Rate (MB/sec)** | 0.001 ± 0.014 | 0.001 ± 0.014 | **Same** |
| **GC Count** | ≈ 0 | ≈ 0 | **Same** |
--> Essentially identical performance
## 2. getWithNullableUuidHolder
Read into `NullableUuidHolder` (supports null values)
| Metric | Old (buffer+offset) | New (two-longs) | Change |
|--------|---------------------|-----------------|--------|
| **Time (µs/op)** | 9.918 ± 0.120 | 9.979 ± 0.621 | ≈ **+0.6%** (0.061 µs
slower) |
| **Memory Allocation (B/op)** | 0.011 ± 0.144 | 0.011 ± 0.144 | **Same** |
| **Allocation Rate (MB/sec)** | 0.001 ± 0.014 | 0.001 ± 0.014 | **Same** |
| **GC Count** | ≈ 0 | ≈ 0 | **Same** |
--> Essentially identical performance, both allocation-free
## 3. getWithUuidHolder
Read into `UuidHolder` (non-nullable)
| Metric | Old (buffer+offset) | New (two-longs) | Change |
|--------|---------------------|-----------------|--------|
| **Time (µs/op)** | 9.934 ± 0.162 | 16.488 ± 0.121 | ⚠️ **+66.0%** (6.554
µs slower) |
| **Memory Allocation (B/op)** | 0.011 ± 0.144 | 0.019 ± 0.239 | **+73%**
(still ~0 bytes) |
| **Allocation Rate (MB/sec)** | 0.001 ± 0.014 | 0.001 ± 0.014 | **Same** |
| **GC Count** | ≈ 0 | ≈ 0 | **Same** |
--> 66% performance regression
- Old approach: Direct buffer reference (10 µs/op)
- New approach: Read two longs + byte reversal (16 µs/op)
## 4. setUuidDirectly
Write UUID objects directly via `vector.setSafe(i, uuid)`
| Metric | Old (buffer+offset) | New (two-longs) | Change |
|--------|---------------------|-----------------|--------|
| **Time (µs/op)** | 87.047 ± 2.229 | 89.154 ± 5.116 | ≈ **+2.4%** (2.107 µs
slower) |
| **Memory Allocation (B/op)** | 320000.101 ± 1.262 | 320000.103 ± 1.290 |
**Same** |
| **Allocation Rate (MB/sec)** | 3505.783 ± 88.171 | 3422.955 ± 195.204 |
**-2.4%** |
| **GC Count** | 174 | 170 | **-2.3%** |
| **GC Time (ms)** | 79 | 78 | **-1.3%** |
--> Essentially identical performance
- High allocation is expected (creating 10,000 UUID objects = 32 bytes each)
## 5. setWithNullableUuidHolder
Write using `NullableUuidHolder` with nullability support
| Metric | Old (buffer+offset) | New (two-longs) | Change |
|--------|---------------------|-----------------|--------|
| **Time (µs/op)** | 25.415 ± 4.776 | 25.052 ± 1.187 | **-1.4%** (0.363 µs
faster) |
| **Memory Allocation (B/op)** | 0.029 ± 0.364 | 0.029 ± 0.362 | **Same** |
| **Allocation Rate (MB/sec)** | 0.001 ± 0.014 | 0.001 ± 0.014 | **Same** |
| **GC Count** | ≈ 0 | ≈ 0 | **Same** |
--> Essentially identical performance
## 6. setWithUuidHolder
Write using `UuidHolder` (non-nullable)
| Metric | Old (buffer+offset) | New (two-longs) | Change |
|--------|---------------------|-----------------|--------|
| **Time (µs/op)** | **FAILED** | 91.831 ± 7.582 | **FIXED** |
| **Memory Allocation (B/op)** | **N/A** | 320000.106 ± 1.328 | **Works** |
| **Allocation Rate (MB/sec)** | **N/A** | 3323.156 ± 272.962 | **Works** |
| **GC Count** | **N/A** | 165 | **Works** |
| **GC Time (ms)** | **N/A** | 75 | **Works** |
**Error in Old Version**:
```
java.lang.NullPointerException: Cannot set UUID vector, the source buffer is
null.
at org.apache.arrow.vector.UuidVector.set(UuidVector.java:241)
```
--> Offset implementation was broken!
- The old buffer+offset approach had a NullPointerException bug
- The new two-longs approach fixes this bug and works correctly
## 7. setWithWriter
Write using `UuidWriterImpl` (extension writer pattern)
| Metric | Old (buffer+offset) | New (two-longs) | Change |
|--------|---------------------|-----------------|--------|
| **Time (µs/op)** | 62.174 ± 8.932 | 61.747 ± 3.117 | **-0.7%** (0.427 µs
faster) |
| **Memory Allocation (B/op)** | 213312.072 ± 0.912 | 213312.076 ± 1.036 |
**Same** |
| **Allocation Rate (MB/sec)** | 3272.006 ± 473.174 | 3294.462 ± 165.505 |
**+0.7%** |
| **GC Count** | 162 | 164 | ≈ **+1.2%** |
| **GC Time (ms)** | 73 | 76 | ≈ **+4.1%** |
--> Essentially identical performance
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
